{"schema_version":2,"benchmark":"JevBench","revision":"v1.8.16","latest_release":"v1.8.16","latest_release_page":"/jev-models/v1.8.16","base_release":"v1.6.1","restated_by":"v1.8.16","method":{"label":"v1.7","decision_limit_s":10,"treatment":"share","rule":"A decision over 10 s (measured latency) counts as a wrong answer and as zero calibration for that item."},"source":{"page":"/jev-models","version_page":"/jev-models/v1.6.1","frozen_page":"/jev-models/v1.6.1","artifact":"/api/jevbench/v1.6.1","artifact_sha256":"5cd8c1332226ea9c179883d2285c0d8b8cc593ae38f2645197364a7ddd9c83c1"},"boards":{"open":"/jev-models","api":"/jev-models/api"},"headline":"capability","composite_option":"A","weights":{"intelligence":25,"calibration":25,"speed":25,"cost":25},"capability_policy":{"definition":"Arithmetic mean of Intelligence and Calibration; ranked models within both official caps","reference_key":"jev-1.13.0@v1.5.7","cost_cap_usd_per_1000":0.06459465517241379,"median_latency_cap_s":1.2329566404223442,"factor":2},"systems":[{"key":"ajev-gemma4-31b-lm6q","name":"AJev lm6q (Gemma 4 31B)","board":"open","source_url":"https://huggingface.co/andyzhang232/ajev-gemma4-31b-lm6q","axes":{"intelligence":71.46073788705554,"calibration":91.33817262545804,"speed":82.45454940142191,"cost":44.17643930820148},"capability":{"score":81.3994552562568,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.2× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":52.29743457752002,"rank":140,"open_board_rank":110,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0725747179002238,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 8583.03975179 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 1,530,069 input + 0 output tokens over 1,485 decisions = 1030.34949495 tokens/decision. GPU-time cost USD 0.0725747179002238/1,000; old list-price cost USD 0.0927314545454545/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0725747179002238/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.21341930099998763,"p50_s_adjusted":0.5768386019999753,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.21341930099998763,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.4175618237000378,"p95_s_adjusted":0.9851236474000756,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"aleria-judge-1.1.1","name":"Aleria Judge 1.1.1 (ALERIA-ML, inference technique on stock Gemma 4 31B, author server, self-hosted)","board":"open","source_url":"https://github.com/ALERIA-ML/aleria-judge","axes":{"intelligence":68.01802916815697,"calibration":85.70385668097765,"speed":86.37984483839423,"cost":41.14441351959124},"capability":{"score":76.86094292456731,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.8× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":43.51029656636035,"rank":160,"open_board_rank":127,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.09159101728590062,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,273,730 input + 0 output tokens over 1,485 decisions = 1531.13131313 tokens/decision. GPU-time cost USD 0.0915910172859006/1,000; old list-price cost USD 0.137801818181818/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0915910172859006/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13238183599992226,"p50_s_adjusted":0.41476367199984454,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13238183599992226,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.202450051499994,"p95_s_adjusted":0.554900102999988,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"allec-oavg","name":"allec-oavg (plantwaterer; commercially restricted)","board":"open","source_url":"https://huggingface.co/plantwaterer/allec","axes":{"intelligence":70.50999328551342,"calibration":91.08404265365112,"speed":84.95032191909448,"cost":55.89119947180174},"capability":{"score":80.79701796958227,"rank":33,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":25,"api_board_rank":null},"composite_score":72.95975097428432,"rank":62,"open_board_rank":47,"api_board_rank":null,"measured_in":"v1.8.3","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.3","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029532033400877865,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 735,102 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0295320334008779/1,000; old list-price estimate USD 0.0444319543317663/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0295320334008779/1,000. Estimated, not a bill. USD 0.0295320334 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0295061687 over all 1,500 items)."},"latency":{"p50_s_raw":0.14198202400007176,"p50_s_adjusted":0.43396404800014354,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.14198202400007176,"latency_basis":"api-end-to-end","p95_s_raw":0.29353951570006415,"p95_s_adjusted":0.7370790314001283,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"amazon-strands-decider-26b-a4b","name":"Amazon Strands Decider 26B-A4B","board":"open","source_url":"https://huggingface.co/amazon/strands-decider-26B-A4B-gemma4-v1-2610","axes":{"intelligence":46.72626896403243,"calibration":86.70498962717834,"speed":87.9627675790118,"cost":71.89605102495447},"capability":{"score":66.71562929560538,"rank":124,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":97,"api_board_rank":null},"composite_score":60.010819075526705,"rank":130,"open_board_rank":100,"api_board_rank":null,"measured_in":"v1.7.1","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.1","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008645662921176805,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 734,625 input + 1,485 output tokens over 1,485 decisions = 495.696969697 tokens/decision. GPU-time cost USD 0.00864566292117681/1,000; old list-price cost USD 0.0448227272727273/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00864566292117681/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11055840500000613,"p50_s_adjusted":0.3711168100000123,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11055840500000613,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.14036871655000027,"p95_s_adjusted":0.43073743310000057,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"amber-4b","name":"Amber 4B (MLukauskas)","board":"open","source_url":"https://huggingface.co/MLukauskas/amber-4b","axes":{"intelligence":24.765045191680287,"calibration":82.75043429348688,"speed":93.44090583444607,"cost":86.42842132051366},"capability":{"score":53.75773974258358,"rank":209,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":175,"api_board_rank":null},"composite_score":13.130068015470234,"rank":241,"open_board_rank":194,"api_board_rank":null,"measured_in":"v1.8.2","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.2","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0028339093623573133,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 429,428 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00283390936235731/1,000; old list-price estimate USD 0.00865200805910007/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00283390936235731/1,000. Estimated, not a bill. Reference deepinfra:Qwen/Qwen3.5-4B is frozen at the 25 Sep cut-off and was deprecated afterwards (replaced by Qwen/Qwen3.5-9B); re-score if it changes (I-3). USD 0.00283390936 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00541156958 over all 1,500 items)."},"latency":{"p50_s_raw":0.026130665000039244,"p50_s_adjusted":0.20226133000007848,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.026130665000039244,"latency_basis":"api-end-to-end","p95_s_raw":0.036935169749983746,"p95_s_adjusted":0.22387033949996749,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.2","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"amos-1-large","name":"Amos 1 large (Scira)","board":"api","source_url":"https://huggingface.co/mukaddamzaid/amos-1-large","axes":{"intelligence":68.17020573384414,"calibration":87.80739845794042,"speed":83.27058287283708,"cost":30.04968825138303},"capability":{"score":77.98880209589228,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 6.6× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":20.25053358459254,"rank":214,"open_board_rank":null,"api_board_rank":43,"measured_in":"v1.6.17","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.21462339393939392,"basis":"ESTIMATE (*est.), exact base-model market reference from the frozen 25 Sep 2026 snapshot (pricing_v15 sha256 dae86740...): qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output, x the provider-reported input tokens; free keyless endpoint, nothing generated (output 0). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.214623 per 1,000 decisions on the common basis (USD 0.542843 over all 1,500 items) is 3.323x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 precedent) and falls outside the official-caps board views by construction. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.214623 per 1,000 decisions (3.323x the official cost cap USD 0.064595), against USD 0.542843 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.532363753503887,"p50_s_adjusted":0.532363753503887,"adjustment":"none (API)","p50_s_measured":0.532363753503887,"latency_basis":"api-end-to-end","p95_s_raw":0.8845720937955772,"p95_s_adjusted":0.8845720937955772,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":21,"n_over_limit_answered_ok":21,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.401636755316044,"capability":79.35913670315466,"axes":{"calibration":88.99486051802106,"cost":30.04968825138303,"intelligence":69.72341288828827,"speed":83.61644312651299},"rank":130,"open_board_rank":null,"api_board_rank":35}},{"key":"amos-1-medium","name":"Amos 1 medium (Scira)","board":"api","source_url":"https://huggingface.co/mukaddamzaid/amos-1-medium","axes":{"intelligence":38.09902514689999,"calibration":87.60140992244203,"speed":86.2337589756876,"cost":48.74716696332005},"capability":{"score":62.85021753467101,"rank":149,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":31},"composite_score":31.638672185087426,"rank":185,"open_board_rank":null,"api_board_rank":39,"measured_in":"v1.6.17","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.05110080808080808,"basis":"ESTIMATE (*est.), exact base-model market reference from the frozen 25 Sep 2026 snapshot (pricing_v15 sha256 dae86740...): qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output, x the provider-reported input tokens; free keyless endpoint, nothing generated (output 0). INSIDE THE OFFICIAL COST CAP on the common basis: USD 0.051101 per 1,000 decisions is 0.791x the official cap USD 0.064595. Over all 1,500 items the row cost USD 0.129248 (2.001x the cap) and was disclosed as ranked over the cap in v1.6.29; that disclosure is superseded. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.051101 per 1,000 decisions (0.791x the official cost cap USD 0.064595), against USD 0.129248 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.44756228550249944,"p50_s_adjusted":0.44756228550249944,"adjustment":"none (API)","p50_s_measured":0.44756228550249944,"latency_basis":"api-end-to-end","p95_s_raw":0.5318272908531071,"p95_s_adjusted":0.5318272908531071,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":10,"n_over_limit_answered_ok":10,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":32.20168753691601,"capability":63.175111580896406,"axes":{"calibration":87.98659582926916,"cost":48.74716696332005,"intelligence":38.36362733252365,"speed":86.48270625779318},"rank":96,"open_board_rank":null,"api_board_rank":31}},{"key":"aplomb-1","name":"Aplomb 1 (5.3B decision model on Qwen3.5-4B, trained readout head, self-hosted)","board":"open","source_url":"https://huggingface.co/empiriolabsai/aplomb-1","axes":{"intelligence":45.23340779864823,"calibration":88.94185474430773,"speed":88.75158098066062,"cost":72.4070134176205},"capability":{"score":67.08763127147799,"rank":122,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":95,"api_board_rank":null},"composite_score":56.028575263593325,"rank":137,"open_board_rank":107,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008313161534050324,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,249,558 input + 0 output tokens over 1,477 decisions = 846.010832769 tokens/decision. GPU-time cost USD 0.00831316153405032/1,000; old list-price cost USD 0.0253803249830738/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00831316153405032/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.09257470816373825,"p50_s_adjusted":0.3351494163274765,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09257470816373825,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12387190728913991,"p95_s_adjusted":0.3977438145782798,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":54.55838408944504,"capability":67.51828279493242,"axes":{"calibration":89.21801638247211,"cost":57.86508463809988,"intelligence":45.81854920739271,"speed":88.88310989631262},"rank":61,"open_board_rank":37,"api_board_rank":null}},{"key":"apus-openjev-v1-35b-a3b","name":"APUS-OpenJev-v1-35B-A3B (merged bf16 checkpoint-5949 MoE on Qwen3.5-35B-A3B, their own native runtime, full depth, self-hosted)","board":"open","source_url":"https://huggingface.co/apus-ailab/APUS-OpenJev-v1-35B-A3B","axes":{"intelligence":52.808647095759525,"calibration":67.91594971090585,"speed":80.338990145866,"cost":null},"capability":{"score":60.36229840333269,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":false,"listing":"listed","not_ranked_because":"registry ranked=false or no composite","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED, and deliberately not invented. pricing_v15.reference() RAISES on the exact declared base Qwen/Qwen3.5-35B-A3B ('base model lacks an explicit market-listing decision'). Run 46's bongard-mini rule applies unchanged: the frozen module raises rather than estimating, and the frozen snapshot requires an explicit base-model listing under the published estimate rule. Complete but unpriced - shown, not ranked. PRICING REFERENCE NOTE: the dated snapshot ALREADY lists the exact base as qwen/qwen3.5-35b-a3b at (0.3125, 1.25), and BASE_REFERENCES ALREADY carries 'Qwen/Qwen3.5-35B-A3B-Base' -> 'qwen/qwen3.5-35b-a3b'. Only the non--Base (instruct) spelling of the same model is missing, so every Qwen3.5-35B-A3B derivative raises today. BECAUSE OUTPUT TOKENS ARE 0 BY CONSTRUCTION this row's Cost depends on the input rate alone, so the number is one multiplication over tokens the raw keeps; r56/receipts/COST-ARITHMETIC-R56.json carries it already computed."},"latency":{"p50_s_raw":0.35987919551553205,"p50_s_adjusted":0.8697583910310641,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.35987919551553205,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.4567069810640532,"p95_s_adjusted":1.0634139621281065,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":60.36229840333269,"axes":{"calibration":67.91594971090585,"cost":null,"intelligence":52.808647095759525,"speed":80.50610262660216},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"apus-openjev-v1-9b","name":"APUS-OpenJev-v1-9B (merged bf16 checkpoint-3000 on Qwen3.5-9B, their own native runtime, full depth, self-hosted)","board":"open","source_url":"https://huggingface.co/apus-ailab/APUS-OpenJev-v1-9B","axes":{"intelligence":47.20514429531085,"calibration":66.61054590286017,"speed":82.0651268880946,"cost":68.40093916429564},"capability":{"score":56.907845099085506,"rank":188,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":154,"api_board_rank":null},"composite_score":56.59069443023599,"rank":136,"open_board_rank":106,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011305818986853587,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 755,283 input + 0 output tokens over 1,477 decisions = 511.362897766 tokens/decision. GPU-time cost USD 0.0113058189868536/1,000; old list-price cost USD 0.0511362897765741/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0113058189868536/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.28152690053684637,"p50_s_adjusted":0.7130538010736928,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.28152690053684637,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3608479586080648,"p95_s_adjusted":0.8716959172161296,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":49.17976892452783,"capability":56.907845099085506,"axes":{"calibration":66.61054590286017,"cost":48.738123569705756,"intelligence":47.20514429531085,"speed":82.18245466956745},"rank":65,"open_board_rank":41,"api_board_rank":null}},{"key":"arcbright-elf-1-12b","name":"Arcbright Elf 1 12B","board":"open","source_url":"https://huggingface.co/arcbright/elf-1-12b","axes":{"intelligence":43.75476217706713,"calibration":86.16803253832052,"speed":91.60645552542691,"cost":67.83160602523185},"capability":{"score":64.96139735769383,"rank":132,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":104,"api_board_rank":null},"composite_score":50.95211596400614,"rank":142,"open_board_rank":111,"api_board_rank":null,"measured_in":"v1.6.34","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-20","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011810813156563906,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 777,809 input + 1,485 output tokens over 1,485 decisions = 524.777104377 tokens/decision. GPU-time cost USD 0.0118108131565639/1,000; old list-price cost USD 0.0263388552188552/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0118108131565639/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.042285979000013185,"p50_s_adjusted":0.23457195800002636,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.042285979000013185,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07224764539998886,"p95_s_adjusted":0.2944952907999777,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":48.80851298722302,"capability":64.96139735769383,"axes":{"calibration":86.16803253832052,"cost":57.38209314881071,"intelligence":43.75476217706713,"speed":91.97994026431576},"rank":66,"open_board_rank":42,"api_board_rank":null}},{"key":"arcbright-elf-1-12b-r2","name":"Arcbright Elf 1 12B (r2)","board":"open","source_url":"https://huggingface.co/arcbright/elf-1-12b","axes":{"intelligence":63.91041297511882,"calibration":91.06809080061666,"speed":92.52857247879322,"cost":71.73785397462883},"capability":{"score":77.48925188786774,"rank":51,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":38,"api_board_rank":null},"composite_score":77.85916639726614,"rank":7,"open_board_rank":4,"api_board_rank":null,"measured_in":"v1.7.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.14","measured_on":"2026-10-27","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008751279087675506,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 577,488 input + 1,489 output tokens over 1,489 decisions = 388.836131632 tokens/decision. GPU-time cost USD 0.00875127908767551/1,000; old list-price cost USD 0.0195418065815984/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00875127908767551/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.028828452000027482,"p50_s_adjusted":0.20765690400005496,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.028828452000027482,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05951365754997937,"p95_s_adjusted":0.26902731509995875,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.14","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"arcbright-elf-1-31b","name":"Arcbright Elf 1 31B","board":"open","source_url":"https://huggingface.co/arcbright/elf-1-31b","axes":{"intelligence":71.94618834920504,"calibration":91.63064878963459,"speed":89.86892425290594,"cost":55.0955403040988},"capability":{"score":81.78841856941982,"rank":22,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":16,"api_board_rank":null},"composite_score":73.95046410857292,"rank":44,"open_board_rank":31,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.031391735265312345,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 777,809 input + 1,485 output tokens over 1,485 decisions = 524.777104377 tokens/decision. GPU-time cost USD 0.0313917352653123/1,000; old list-price cost USD 0.0474799393939394/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0313917352653123/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06162827900004686,"p50_s_adjusted":0.27325655800009374,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06162827900004686,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11358493024998212,"p95_s_adjusted":0.37716986049996426,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"arsenelupin-v1.1","name":"ArseneLupin v1.1 (IBOU-SEARCH, Qwen3.5-4B LoRA)","board":"open","source_url":"https://huggingface.co/IBOU-SEARCH/ArseneLupin-v1.1","axes":{"intelligence":27.956374646954437,"calibration":82.22627427476863,"speed":89.52285293002008,"cost":75.23279561612854},"capability":{"score":55.09132446086153,"rank":200,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":166,"api_board_rank":null},"composite_score":17.27345398717365,"rank":223,"open_board_rank":179,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006692270388071554,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,011,369 input + 0 output tokens over 1,485 decisions = 681.056565657 tokens/decision. GPU-time cost USD 0.00669227038807155/1,000; old list-price cost USD 0.020431696969697/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00669227038807155/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0849640910018934,"p50_s_adjusted":0.3199281820037868,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0849640910018934,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09943443111260418,"p95_s_adjusted":0.34886886222520835,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.550252807567574,"capability":55.09132446086153,"axes":{"calibration":82.22627427476863,"cost":60.69086683660791,"intelligence":27.956374646954437,"speed":89.69506927060962},"rank":151,"open_board_rank":114,"api_board_rank":null}},{"key":"attest-31b-v1.0","name":"attest-31B-v1.0","board":"api","source_url":"https://attest.engineering/model","axes":{"intelligence":65.23849472031306,"calibration":73.84793106824938,"speed":96.47488998719595,"cost":50.72941039177032},"capability":{"score":69.54321289428123,"rank":109,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":26},"composite_score":67.85694940132115,"rank":115,"open_board_rank":null,"api_board_rank":25,"measured_in":"v1.6.13","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.04388872727272727,"basis":"ESTIMATE, v1.5-M2 base-model reference rule: vendor-hosted API, customer list price USD 0.048/M input and USD 0.00/M output effective 2026-10-11 has been in force fewer than 30 days. Base model google/gemma-4-31b-it reference on OpenRouter USD 0.09/M input / USD 0.34/M output, snapshot fetched 2026-10-12T17:55:05Z (SHA-256 81eae2f3f94bf299c8d4499260f842c357b7a869ce82dfa1741ad6a89cf57014), same base-model estimate precedent as metask_jev_rain_31b_trained. Applied to measured response usage; failed HTTP 500 usage/billing may be unreported. No output generation; reported output tokens 0. Estimated, not a vendor bill. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.043889 per 1,000 decisions (0.679x the official cost cap USD 0.064595), against USD 0.043450 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.11031594549967849,"p50_s_adjusted":0.11031594549967849,"adjustment":"none (API)","p50_s_measured":0.11031594549967849,"latency_basis":"api-end-to-end","p95_s_raw":0.20411396054878422,"p95_s_adjusted":0.20411396054878422,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.9350679784662,"capability":69.54321289428123,"axes":{"calibration":73.84793106824938,"cost":50.72941039177032,"intelligence":65.23849472031306,"speed":97.10993413996468},"rank":38,"open_board_rank":null,"api_board_rank":20}},{"key":"autojev-27b","name":"AutoJev-27B (denis-pplx, Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/denis-pplx/autojev-27b","axes":{"intelligence":55.62422706919757,"calibration":87.61544276592178,"speed":88.30997360530876,"cost":56.735701109968524},"capability":{"score":71.61983491755967,"rank":86,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":63,"api_board_rank":null},"composite_score":68.56285633198503,"rank":111,"open_board_rank":87,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02767854753983507,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 874,592 input + 0 output tokens over 1,624 decisions = 538.541871921 tokens/decision. GPU-time cost USD 0.0276785475398351/1,000; old list-price cost USD 0.226187586206897/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0276785475398351/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10048181959427893,"p50_s_adjusted":0.3509636391885579,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10048181959427893,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.13523766243597493,"p95_s_adjusted":0.42047532487194983,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.463427917707932,"capability":71.61983491755967,"axes":{"calibration":87.61544276592178,"cost":29.365937020749882,"intelligence":55.62422706919757,"speed":88.60757446254297},"rank":140,"open_board_rank":105,"api_board_rank":null}},{"key":"basal-omni-1.0-ultra","name":"basal-omni-1.0-ultra (Remek Kinas, Gemma-4-31B-it + merged LoRA, typed-decision readout, self-hosted)","board":"open","source_url":"https://huggingface.co/Remek/basal-omni-1.0-ultra","axes":{"intelligence":74.05810874808022,"calibration":90.20661225231586,"speed":86.63193265367512,"cost":55.924617993700764},"capability":{"score":82.13236050019805,"rank":18,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":13,"api_board_rank":null},"composite_score":74.05642013689533,"rank":41,"open_board_rank":28,"api_board_rank":null,"measured_in":"v1.6.28","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-17","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029456381792491746,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10285.272649 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 744,183 input + 0 output tokens over 1,485 decisions = 501.133333333 tokens/decision. GPU-time cost USD 0.0294563817924917/1,000; old list-price cost USD 0.045102/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0294563817924917/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10769466249621473,"p50_s_adjusted":0.3653893249924295,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10769466249621473,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2221808933332795,"p95_s_adjusted":0.594361786666559,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":71.69146765135073,"capability":82.53799311737751,"axes":{"calibration":91.0178774866748,"cost":50.37412598110784,"intelligence":74.05810874808022,"speed":87.31015625661291},"rank":17,"open_board_rank":6,"api_board_rank":null}},{"key":"bb-qwen3.5-4b-lora","name":"BB-Qwen3.5-4B-LoRA (Babak Barazandeh, LoRA on Qwen3.5-4B-Base, hosted /v1/systemone)","board":"api","source_url":"https://github.com/fstandhartinger/jevbench/issues/156","axes":{"intelligence":38.67355961154223,"calibration":82.13857452060611,"speed":80.91147986788948,"cost":null},"capability":{"score":60.40606706607417,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":false,"listing":"listed","not_ranked_because":"registry ranked=false or no composite","price":{"kind":"estimated","usd_per_1000_decisions":null,"basis":"UNPRICED - NOT A MEASUREMENT FAILURE. The base is disclosed (Qwen/Qwen3.5-4B-Base; reference() resolves it to deepinfra:Qwen/Qwen3.5-4B at 0.03/0.15), but the endpoint reports usage.input_tokens ONLY - no output_tokens on any answer - and the system is API-only, so 'nothing is generated' cannot be verified in its code; the published 02 Oct rule keeps an unverified null unpriced. The frozen scorer therefore gives no Cost axis and no composite; the row is reported on Capability. PRICING REFERENCE NOTE: if the author confirms (or independently verified evidence establishes) a one-pass readout with 0 output tokens, the row prices by one multiplication from the raw at 0.03/M input (receipts/COST-ALTERNATIVES-R59.json)."},"latency":{"p50_s_raw":0.7166377417743206,"p50_s_adjusted":0.7166377417743206,"adjustment":"none (API)","p50_s_measured":0.7166377417743206,"latency_basis":"api-end-to-end","p95_s_raw":1.131233705580234,"p95_s_adjusted":1.131233705580234,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":60.40606706607417,"axes":{"calibration":82.13857452060611,"cost":null,"intelligence":38.67355961154223,"speed":81.57813729165589},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"bespoke-nimble-9b-v3","name":"Bespoke Nimble 9B v3 (Qwen3.5-9B LoRA)","board":"open","source_url":"https://huggingface.co/bespokelabs/Bespoke-Nimble-9B-v3","axes":{"intelligence":56.54011796945148,"calibration":77.04569964036838,"speed":89.00367576375322,"cost":65.75776727887356},"capability":{"score":66.79290880490993,"rank":123,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":96,"api_board_rank":null},"composite_score":70.04188420031245,"rank":98,"open_board_rank":76,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.013848660139868772,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 925,157 input + 0 output tokens over 1,477 decisions = 626.375761679 tokens/decision. GPU-time cost USD 0.0138486601398688/1,000; old list-price cost USD 0.0626375761679079/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0138486601398688/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.09450600598938763,"p50_s_adjusted":0.3390120119787753,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09450600598938763,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11051855008554409,"p95_s_adjusted":0.37103710017108815,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":53.47656993239491,"capability":66.79290880490993,"axes":{"calibration":77.04569964036838,"cost":46.094951684283686,"intelligence":56.54011796945148,"speed":89.19502471208428},"rank":63,"open_board_rank":39,"api_board_rank":null}},{"key":"bev-bonsai-27b","name":"Bev / Bonsai 27B","board":"open","source_url":"https://huggingface.co/Reza2kn/Bev","axes":{"intelligence":43.77513601921869,"calibration":88.50232899798883,"speed":72.92528840336786,"cost":55.738539603831974},"capability":{"score":66.13873250860377,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.5× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":46.598366686554485,"rank":147,"open_board_rank":115,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02988009770220635,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 942,533 input + 1,624 output tokens over 1,624 decisions = 581.377463054 tokens/decision. GPU-time cost USD 0.0298800977022063/1,000; old list-price cost USD 0.246758534482759/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0298800977022063/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.9887027884833515,"p50_s_adjusted":2.127405576966703,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.9887027884833515,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.1233698406023904,"p95_s_adjusted":2.3967396812047808,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":11.7393704396132,"capability":66.13873250860377,"axes":{"calibration":88.50232899798883,"cost":28.231834531826323,"intelligence":43.77513601921869,"speed":72.98222461463935},"rank":164,"open_board_rank":125,"api_board_rank":null}},{"key":"bge-reranker-v2-m3","name":"BAAI bge-reranker-v2-m3","board":"open","source_url":"https://huggingface.co/BAAI/bge-reranker-v2-m3","axes":{"intelligence":0.06512222845080358,"calibration":88.10046932372114,"speed":93.5678736531519,"cost":59.325605115070715},"capability":{"score":44.08279577608597,"rank":235,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":198,"api_board_rank":null},"composite_score":4.4076704742616216e-7,"rank":311,"open_board_rank":258,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.022688885467980296,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: base-model market reference (deepinfra:BAAI/bge-m3)"},"latency":{"p50_s_raw":0.018346461933106184,"p50_s_adjusted":0.18669292386621236,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.018346461933106184,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.042775452532805486,"p95_s_adjusted":0.23555090506561097,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":4.4076803151681546e-7,"capability":44.08279577608597,"axes":{"calibration":88.10046932372114,"cost":59.325605115070715,"intelligence":0.06512222845080358,"speed":93.86976236118387},"rank":238,"open_board_rank":194,"api_board_rank":null}},{"key":"blink-4b","name":"blink-4b (thegovind, Qwen3.5-4B derivative)","board":"open","source_url":"https://huggingface.co/thegovind/blink-4b","axes":{"intelligence":35.035615417279956,"calibration":87.79893140350136,"speed":91.21351546851116,"cost":78.06391380283692},"capability":{"score":61.417273410390656,"rank":164,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":132,"api_board_rank":null},"composite_score":30.829294674656747,"rank":189,"open_board_rank":150,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005385212847761947,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 813,840 input + 0 output tokens over 1,485 decisions = 548.04040404 tokens/decision. GPU-time cost USD 0.00538521284776195/1,000; old list-price cost USD 0.0164412121212121/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00538521284776195/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05763030983507633,"p50_s_adjusted":0.2652606196701527,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05763030983507633,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06754292871337383,"p95_s_adjusted":0.28508585742674764,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":30.012225474173363,"capability":61.62516934780887,"axes":{"calibration":87.95885156098548,"cost":63.5219850233163,"intelligence":35.29148713463226,"speed":91.30713992573452},"rank":101,"open_board_rank":69,"api_board_rank":null}},{"key":"blink-v0-3-26b-a4b-nvfp4","name":"Blink v0.3 26B-A4B","board":"open","source_url":"https://huggingface.co/PixilabAI/Blink-v0.3-26B-A4B-NVFP4","axes":{"intelligence":49.21581563112197,"calibration":90.67498392881862,"speed":91.13697610522254,"cost":71.06457524185485},"capability":{"score":69.9453997799703,"rank":107,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":81,"api_board_rank":null},"composite_score":68.7254259320887,"rank":108,"open_board_rank":84,"api_board_rank":null,"measured_in":"v1.6.7","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00921539931730195,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 791,073 input + 1,471 output tokens over 1,500 decisions = 528.362666667 tokens/decision. GPU-time cost USD 0.00921539931730195/1,000; old list-price cost USD 0.04775858/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00921539931730195/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05238986399490386,"p50_s_adjusted":0.25477972798980775,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05238986399490386,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07604540715692563,"p95_s_adjusted":0.30209081431385126,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":61.16059214787521,"capability":69.9453997799703,"axes":{"calibration":90.67498392881862,"cost":49.62845783533821,"intelligence":49.21581563112197,"speed":91.48151024194223},"rank":57,"open_board_rank":33,"api_board_rank":null}},{"key":"bobcat-1.3","name":"Bobcat 1.3 (Gemma 4 31B)","board":"open","source_url":"https://huggingface.co/sanghwa-na/bobcat-1.3","axes":{"intelligence":68.0935397642001,"calibration":87.62628477134197,"speed":89.29862693377845,"cost":55.19130539235455},"capability":{"score":77.85991226777104,"rank":48,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":35,"api_board_rank":null},"composite_score":72.1826764470016,"rank":73,"open_board_rank":56,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.031161844321510472,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 773,587 input + 0 output tokens over 1,485 decisions = 520.934006734 tokens/decision. GPU-time cost USD 0.0311618443215105/1,000; old list-price cost USD 0.0468840606060606/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0311618443215105/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06891351249998934,"p50_s_adjusted":0.2878270249999787,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06891351249998934,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12916239049991987,"p95_s_adjusted":0.4083247809998397,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"bobcat-flash-1.2","name":"Bobcat Flash 1.2 (Gemma-4-26B-A4B-it + two merged rank-64 LoRA adapters, typed-decision readout at the first answer position, FP8, self-hosted)","board":"open","source_url":"https://huggingface.co/sanghwa-na/bobcat-flash-1.2","axes":{"intelligence":65.56939608070681,"calibration":80.53028652577869,"speed":90.60003447236298,"cost":71.06558799286785},"capability":{"score":73.04984130324274,"rank":79,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":56,"api_board_rank":null},"composite_score":75.7895318580443,"rank":20,"open_board_rank":13,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009214683018211383,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 780,331 input + 0 output tokens over 1,477 decisions = 528.321597833 tokens/decision. GPU-time cost USD 0.00921468301821138/1,000; old list-price cost USD 0.0475489438050102/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00921468301821138/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.061533288506325334,"p50_s_adjusted":0.2730665770126507,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.061533288506325334,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08447698290110565,"p95_s_adjusted":0.3189539658022113,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":13,"n_over_limit_answered_ok":13,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.34032787154446,"capability":73.4936722737144,"axes":{"calibration":81.17269878237248,"cost":49.68577376542501,"intelligence":65.81464576505633,"speed":90.781726629608},"rank":43,"open_board_rank":23,"api_board_rank":null}},{"key":"bongard-mini","name":"Bongard-mini (AgentBull, T5Gemma-2 4B-4B)","board":"open","source_url":"https://huggingface.co/AgentBull/bongard-mini","axes":{"intelligence":25.21389259164919,"calibration":84.35778065081391,"speed":90.4627802720504,"cost":null},"capability":{"score":54.78583662123155,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base google/t5gemma-2-4b-4b and raises rather than estimating; one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 dcfcecb0...2835, and every later re-freeze, which only add hosted-API launch-list keys) RAISES on the declared base google/t5gemma-2-4b-4b ('base model lacks an explicit market-listing decision'); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens as reported by the author's server (input 786,538, output_tokens 0)."},"latency":{"p50_s_raw":0.06963305150566157,"p50_s_adjusted":0.28926610301132316,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06963305150566157,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08037975499464664,"p95_s_adjusted":0.31075950998929325,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":54.78583662123155,"axes":{"calibration":84.35778065081391,"cost":null,"intelligence":25.21389259164919,"speed":90.58298788933249},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"celeris-1-decision","name":"Celeris-1 Decision (Celeris)","board":"api","source_url":"https://docs.celeris.ai","axes":{"intelligence":58.890192862872325,"calibration":88.35471243386242,"speed":86.92599216009037,"cost":60.373088134904265},"capability":{"score":73.62245264836737,"rank":74,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":20},"composite_score":70.96415918238667,"rank":88,"open_board_rank":null,"api_board_rank":20,"measured_in":"v1.6.25","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020936161616161617,"basis":"Celeris public price, USD 0.04/M input, output free (celeris.ai/pricing, fetched 16 Oct 2026 19:27Z, saved copy sha256 c3ed0fdc...; docs.celeris.ai/pricing rates table, saved copy sha256 ff4b658d..., byte-identical to the copy saved on 14 Oct 2026): \"Celeris-1 Decision $0.04 / M input tokens, Free output tokens\"; prepaid per-token billing with no subscription or commitment. Counted as the manufacturer's standard launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap, d1-ultra, Sage 1.4, Nimble and kettei), through a scoped re-freeze of the frozen pricing module that adds this one key and nothing else, with a zero-difference proof for every other registry entry. Workspace activation at Celeris is capacity-limited (an activation waitlist, as for SubQ Snap); free or promotional credit is not used. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (1,968,382 input tokens over the 1,500 answered items, 0 output tokens), divided by the 1,485 items of the common cost basis. Celeris counts input tokens its own way: per its docs, input_tokens \"covers everything the model reads for the request, which is more than your state and questions alone\"; on the same 1,500 items it reports about 3.0x the input tokens kettei-fat reports, so the cost reflects Celeris's own count. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.020936 per 1,000 decisions (0.324x the official cost cap USD 0.064595), against USD 0.052490 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.43277494149515405,"p50_s_adjusted":0.43277494149515405,"adjustment":"none (API)","p50_s_measured":0.43277494149515405,"latency_basis":"api-end-to-end","p95_s_raw":0.46896310617739795,"p95_s_adjusted":0.46896310617739795,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.97181314040431,"capability":73.62245264836737,"axes":{"calibration":88.35471243386242,"cost":60.373088134904265,"intelligence":58.890192862872325,"speed":86.97194897947064},"rank":20,"open_board_rank":null,"api_board_rank":13}},{"key":"certo","name":"Certo v1 (AltSlate Labs)","board":"open","source_url":"https://huggingface.co/altslate/certo-decision-model","axes":{"intelligence":0.20278658812856143,"calibration":86.14595667337036,"speed":92.44797254495796,"cost":96.9130801359906},"capability":{"score":43.17437163074946,"rank":240,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":203,"api_board_rank":null},"composite_score":0.000013254504525307007,"rank":309,"open_board_rank":256,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0012673522167487683,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.04227160010486841,"p50_s_adjusted":0.23454320020973682,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04227160010486841,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.046324889175593846,"p95_s_adjusted":0.24264977835118767,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.000013254531668229452,"capability":43.17437163074946,"axes":{"calibration":86.14595667337036,"cost":96.9130801359906,"intelligence":0.20278658812856143,"speed":92.53493469610669},"rank":236,"open_board_rank":192,"api_board_rank":null}},{"key":"chaoliangunsw-jev-style-2b-decision-v4","name":"Jev-Style-2B","board":"open","source_url":"https://huggingface.co/chaoliangUNSW/Jev-Style-2B-Decision-v4","axes":{"intelligence":31.032959958178143,"calibration":86.65104789288367,"speed":94.44817145387303,"cost":90.98733169608725},"capability":{"score":58.8420039255309,"rank":174,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":142,"api_board_rank":null},"composite_score":23.581383853013346,"rank":207,"open_board_rank":166,"api_board_rank":null,"measured_in":"v1.7.17","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.17","measured_on":"2026-10-28","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0019972033083885,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 491,898 input + 0 output tokens over 1,489 decisions = 330.354600403 tokens/decision. GPU-time cost USD 0.0019972033083885/1,000; old list-price cost USD 0.0066070920080591/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0019972033083885/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.013410373999988678,"p50_s_adjusted":0.17682074799997735,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.013410373999988678,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.02653590299998996,"p95_s_adjusted":0.20307180599997993,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.17","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"chon-1.0-12b","name":"Chon-1.0-12B","board":"open","source_url":"https://huggingface.co/ninhnv/chon-1.0-12b","axes":{"intelligence":27.33769023776734,"calibration":58.650234835105636,"speed":64.6591918236347,"cost":64.94195253546884},"capability":{"score":42.993962536436484,"rank":242,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":205,"api_board_rank":null},"composite_score":14.152024957404182,"rank":238,"open_board_rank":191,"api_board_rank":null,"measured_in":"v1.6.32","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-19","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.014743533541759862,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 714,121 input + 258,678 output tokens over 1,485 decisions = 655.083501684 tokens/decision. GPU-time cost USD 0.0147435335417599/1,000; old list-price cost USD 0.050173569023569/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0147435335417599/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1801379819999056,"p50_s_adjusted":0.5102759639998112,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1801379819999056,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":33.4405005668498,"p95_s_adjusted":67.0310011336996,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":540,"n_over_limit_answered_ok":540,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":64.84806528194059,"capability":85.33585692323442,"axes":{"calibration":90.80769377766468,"cost":48.98575014908867,"intelligence":79.86402006880415,"speed":65.53984816070655},"rank":49,"open_board_rank":28,"api_board_rank":null}},{"key":"chon-1.0-12b-fast","name":"Chon-1.0-12B-Fast","board":"open","source_url":"https://huggingface.co/ninhnv/chon-1.0-12b-fast","axes":{"intelligence":60.27202384762607,"calibration":79.86233614365636,"speed":83.73773084289927,"cost":69.09820245715657},"capability":{"score":70.06717999564123,"rank":105,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":79,"api_board_rank":null},"composite_score":72.0367713753772,"rank":75,"open_board_rank":58,"api_board_rank":null,"measured_in":"v1.7.1","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.1","measured_on":"2026-10-25","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01071667149217047,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 707,101 input + 0 output tokens over 1,485 decisions = 476.162289562 tokens/decision. GPU-time cost USD 0.0107166714921705/1,000; old list-price cost USD 0.0238081144781145/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0107166714921705/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.17493440899988855,"p50_s_adjusted":0.4998688179997771,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.17493440899988855,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.34800049334999694,"p95_s_adjusted":0.8460009866999939,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"chon-1.0-e4b","name":"Chon-1.0-E4B","board":"open","source_url":"https://huggingface.co/ninhnv/chon-1.0-e4b","axes":{"intelligence":55.488532895897606,"calibration":81.517233540399,"speed":88.71017618901561,"cost":83.52175108923268},"capability":{"score":68.5028832181483,"rank":118,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":91,"api_board_rank":null},"composite_score":74.71792231346033,"rank":32,"open_board_rank":21,"api_board_rank":null,"measured_in":"v1.7.3","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.3","measured_on":"2026-10-25","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003542215374287284,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 701,161 input + 0 output tokens over 1,485 decisions = 472.162289562 tokens/decision. GPU-time cost USD 0.00354221537428728/1,000; old list-price cost USD 0.00944324579124579/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00354221537428728/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07364347150002004,"p50_s_adjusted":0.2972869430000401,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07364347150002004,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.15134794204997207,"p95_s_adjusted":0.4526958840999441,"n":1500},"method":"v1.7","rescored_in":null,"remeasured_in":"v1.7.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.7.2","method":"v1.6","composite_score":45.93476445894657,"capability":68.39271102371643,"axes":{"calibration":81.82393852673533,"cost":70.74636118712607,"intelligence":54.96148352069754,"speed":43.95627533510597},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"classifier-dev-fast","name":"classifier.dev (fast tier)","board":"api","source_url":"https://classifier.dev","axes":{"intelligence":65.16,"calibration":91.33,"speed":78.21,"cost":58.89},"capability":{"score":78.245,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":71.35,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":false,"listing":"wrapper","not_ranked_because":"Wrapper that serves Jev: listed, never ranked.","price":{"kind":"estimate","usd_per_1000_decisions":0.02345172413793104,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (proxy tokens, I-2): operator list price: higher of 19 Sep plan cost and 26 Sep usage tariff USD 0.042/M input (rule 1.2); no exact base-model floor applies"},"latency":{"p50_s_raw":0.98,"p50_s_adjusted":0.98,"adjustment":"none (API)","p50_s_measured":0.98,"latency_basis":"api-end-to-end","p95_s_raw":1.542,"p95_s_adjusted":1.542,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":71.62,"capability":78.25,"axes":{"calibration":91.33,"cost":58.89,"intelligence":65.16,"speed":79.54},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"classone-gemma4-e2b","name":"ClassOne Gemma 4 E2B (Gemma-4-E2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/devops-thiago/classone-gemma4-e2b","axes":{"intelligence":7.802105373202749,"calibration":22.615652963639242,"speed":93.07921198047904,"cost":85.79078096700756},"capability":{"score":15.208879168420996,"rank":279,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":240,"api_board_rank":null},"composite_score":0.500014420840021,"rank":280,"open_board_rank":228,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.002976052816466239,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 727,075 input + 0 output tokens over 1,477 decisions = 492.264725796 tokens/decision. GPU-time cost USD 0.00297605281646624/1,000; old list-price cost USD 0.00984529451591063/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00297605281646624/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02947877999395132,"p50_s_adjusted":0.20895755998790264,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02947877999395132,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04275808051606873,"p95_s_adjusted":0.23551616103213746,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.49351800738768026,"capability":15.208879168420996,"axes":{"calibration":22.615652963639242,"cost":70.20313863040258,"intelligence":7.802105373202749,"speed":93.2878888255429},"rank":206,"open_board_rank":163,"api_board_rank":null}},{"key":"classone-gemma4-e4b","name":"ClassOne Gemma 4 E4B (Gemma-4-E4B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/devops-thiago/classone-gemma4-e4b","axes":{"intelligence":28.182424563225457,"calibration":48.5204641422835,"speed":91.65069544774573,"cost":82.05537563260856},"capability":{"score":38.35144435275448,"rank":253,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":216,"api_board_rank":null},"composite_score":16.047288987128645,"rank":231,"open_board_rank":185,"api_board_rank":null,"measured_in":"v1.6.9","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003964187114954465,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 792,614 input + 0 output tokens over 1,500 decisions = 528.409333333 tokens/decision. GPU-time cost USD 0.00396418711495447/1,000; old list-price cost USD 0.0105681866666667/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00396418711495447/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0394395335461013,"p50_s_adjusted":0.2288790670922026,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0394395335461013,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07438066378817894,"p95_s_adjusted":0.29876132757635787,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.607101626553092,"capability":38.35144435275448,"axes":{"calibration":48.5204641422835,"cost":69.27998573050196,"intelligence":28.182424563225457,"speed":91.7664962211246},"rank":158,"open_board_rank":120,"api_board_rank":null}},{"key":"classone-qwen3.5-2b","name":"ClassOne Qwen 3.5 2B (Qwen3.5-2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/devops-thiago/classone-qwen3.5-2b","axes":{"intelligence":5.814527660927306,"calibration":32.62324971930184,"speed":92.94182096467813,"cost":85.83730649437784},"capability":{"score":19.21888869011457,"rank":276,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":237,"api_board_rank":null},"composite_score":0.24036918429679566,"rank":285,"open_board_rank":232,"api_board_rank":null,"measured_in":"v1.6.9","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0029654443849206355,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 735,765 input + 0 output tokens over 1,500 decisions = 490.51 tokens/decision. GPU-time cost USD 0.00296544438492064/1,000; old list-price cost USD 0.0098102/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00296544438492064/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03125813900260255,"p50_s_adjusted":0.2125162780052051,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03125813900260255,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0445076500705909,"p95_s_adjusted":0.23901530014118177,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.23766015025378578,"capability":19.21888869011457,"axes":{"calibration":32.62324971930184,"cost":70.24966415777286,"intelligence":5.814527660927306,"speed":93.11279379161611},"rank":212,"open_board_rank":168,"api_board_rank":null}},{"key":"classone-qwen3.5-4b","name":"ClassOne Qwen 3.5 4B (Qwen3.5-4B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/devops-thiago/classone-qwen3.5-4b","axes":{"intelligence":27.338167706853277,"calibration":48.84618651494331,"speed":91.03995916839054,"cost":79.50885516562305},"capability":{"score":38.09217711089829,"rank":254,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":217,"api_board_rank":null},"composite_score":14.833894246342627,"rank":235,"open_board_rank":188,"api_board_rank":null,"measured_in":"v1.6.9","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004819901478944551,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 735,765 input + 0 output tokens over 1,500 decisions = 490.51 tokens/decision. GPU-time cost USD 0.00481990147894455/1,000; old list-price cost USD 0.0147153/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00481990147894455/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04766369954450056,"p50_s_adjusted":0.2453273990890011,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04766369954450056,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08540874201571569,"p95_s_adjusted":0.3208174840314314,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":14.342131800224653,"capability":38.09217711089829,"axes":{"calibration":48.84618651494331,"cost":64.96692638610241,"intelligence":27.338167706853277,"speed":91.46618934008248},"rank":161,"open_board_rank":122,"api_board_rank":null}},{"key":"classone-qwen3.5-9b","name":"ClassOne Qwen 3.5 9B (Qwen3.5-9B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/devops-thiago/classone-qwen3.5-9b","axes":{"intelligence":42.225671849648506,"calibration":58.71667533779978,"speed":89.69784178772274,"cost":68.99811555861743},"capability":{"score":50.471173593724146,"rank":217,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":181,"api_board_rank":null},"composite_score":42.993233043857074,"rank":165,"open_board_rank":131,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01079931348221736,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 721,446 input + 0 output tokens over 1,477 decisions = 488.453622207 tokens/decision. GPU-time cost USD 0.0107993134822174/1,000; old list-price cost USD 0.0488453622207177/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0107993134822174/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06403286599379499,"p50_s_adjusted":0.27806573198759,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06403286599379499,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11776951495121465,"p95_s_adjusted":0.3855390299024293,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":38.54185769307265,"capability":50.471173593724146,"axes":{"calibration":58.71667533779978,"cost":49.33529996402755,"intelligence":42.225671849648506,"speed":90.24509258883624},"rank":85,"open_board_rank":55,"api_board_rank":null}},{"key":"clef","name":"Clef (Cloudflare, Qwen3.8-27B post-train with a joint schema head, multimodal, measured on text)","board":"open","source_url":"https://huggingface.co/Cloudflare/clef","axes":{"intelligence":59.08165501447566,"calibration":91.57703482429629,"speed":82.80534424231135,"cost":55.47192130933339},"capability":{"score":75.32934491938597,"rank":63,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":48,"api_board_rank":null},"composite_score":69.02603466604376,"rank":104,"open_board_rank":81,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.030497854608258067,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 963,677 input + 0 output tokens over 1,624 decisions = 593.397167488 tokens/decision. GPU-time cost USD 0.0304978546082581/1,000; old list-price cost USD 0.249226810344828/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0304978546082581/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.23025226220488548,"p50_s_adjusted":0.610504524409771,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.23025226220488548,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3542859682813287,"p95_s_adjusted":0.8585719365626574,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.753692861922005,"capability":75.32934491938597,"axes":{"calibration":91.57703482429629,"cost":28.102157220114762,"intelligence":59.08165501447566,"speed":83.3998209667487},"rank":150,"open_board_rank":113,"api_board_rank":null}},{"key":"clef-flash","name":"Clef-Flash (Cloudflare, Qwen3.5-9B post-train with a joint schema head, multimodal, measured on text)","board":"open","source_url":"https://huggingface.co/Cloudflare/clef-flash","axes":{"intelligence":42.18368888957246,"calibration":85.46479352633332,"speed":88.14167964175255,"cost":66.46245152664164},"capability":{"score":63.824241207952895,"rank":140,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":110,"api_board_rank":null},"composite_score":46.07172412904015,"rank":148,"open_board_rank":116,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.013119530165836881,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 963,677 input + 0 output tokens over 1,624 decisions = 593.397167488 tokens/decision. GPU-time cost USD 0.0131195301658369/1,000; old list-price cost USD 0.0593397167487685/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0131195301658369/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10368826147168875,"p50_s_adjusted":0.3573765229433775,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10368826147168875,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.13962288135662676,"p95_s_adjusted":0.42924576271325354,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":36.635050479655426,"capability":63.824241207952895,"axes":{"calibration":85.46479352633332,"cost":46.79963593205177,"intelligence":42.18368888957246,"speed":88.4044196893156},"rank":90,"open_board_rank":60,"api_board_rank":null}},{"key":"clef-omni","name":"Clef-omni","board":"open","source_url":"https://huggingface.co/Cloudflare/clef-omni","axes":{"intelligence":34.47716016566848,"calibration":82.15469329918429,"speed":86.57770472711044,"cost":56.02577392355937},"capability":{"score":58.315926732426384,"rank":175,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":143,"api_board_rank":null},"composite_score":26.94791782828204,"rank":199,"open_board_rank":158,"api_board_rank":null,"measured_in":"v1.6.6","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02922856782926389,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,755,730 input + 0 output tokens over 1,500 decisions = 1837.15333333 tokens/decision. GPU-time cost USD 0.0292285678292639/1,000; old list-price cost USD 0.275573/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0292285678292639/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.14697313238866627,"p50_s_adjusted":0.44394626477733257,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.14697313238866627,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.1726675598649308,"p95_s_adjusted":0.49533511972986155,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":6.06640325953113,"capability":58.315926732426384,"axes":{"calibration":82.15469329918429,"cost":26.792900075221908,"intelligence":34.47716016566848,"speed":86.7880661703893},"rank":183,"open_board_rank":143,"api_board_rank":null}},{"key":"clm-8b","name":"CLM-8B (Contrastive-LM, clm-latest)","board":"open","source_url":"https://github.com/Contrastive-LM/CLM","axes":{"intelligence":0.13075619159747265,"calibration":40.03295405037662,"speed":91.7219631344249,"cost":73.74462936099019},"capability":{"score":20.081855120987044,"rank":275,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":236,"api_board_rank":null},"composite_score":0.0000035539240822933435,"rank":310,"open_board_rank":257,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.007502033882766164,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 8B active parameters; throughput 30762.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 619,933 input + 0 output tokens over 1,624 decisions = 381.732142857 tokens/decision. GPU-time cost USD 0.00750203388276616/1,000; old list-price cost USD 0.0446626607142857/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00750203388276616/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03665110794827342,"p50_s_adjusted":0.22330221589654684,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03665110794827342,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0756193115375936,"p95_s_adjusted":0.30123862307518723,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.0000035510969933114796,"capability":20.086328406467768,"axes":{"calibration":40.041900621338065,"cost":50.501662240594015,"intelligence":0.13075619159747265,"speed":92.63548383949004},"rank":237,"open_board_rank":193,"api_board_rank":null}},{"key":"coco-decision-4b","name":"CoCo-Decision-4B (corners-ai, LoRA on Qwen3.5-4B, option-label logits of one pass, served by oh-my-jev)","board":"open","source_url":"https://huggingface.co/corners-ai/CoCo-Decision-4B-Ko","axes":{"intelligence":32.65682750200105,"calibration":81.18327994577871,"speed":91.2364277414064,"cost":79.71332942596933},"capability":{"score":56.92005372388988,"rank":187,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":153,"api_board_rank":null},"composite_score":25.680720840435747,"rank":202,"open_board_rank":161,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004744848524364423,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 713,202 input + 0 output tokens over 1,477 decisions = 482.872037915 tokens/decision. GPU-time cost USD 0.00474484852436442/1,000; old list-price cost USD 0.0144861611374408/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00474484852436442/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0452090189937735,"p50_s_adjusted":0.240418037987547,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0452090189937735,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08144445748621365,"p95_s_adjusted":0.3128889149724273,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":24.64791535308594,"capability":56.92005372388988,"axes":{"calibration":81.18327994577871,"cost":65.17140064644872,"intelligence":32.65682750200105,"speed":91.3615355688411},"rank":113,"open_board_rank":80,"api_board_rank":null}},{"key":"compass-0.2.0","name":"Compass 0.2.0 (adimyth, Qwen3.5-4B LoRA)","board":"open","source_url":"https://huggingface.co/adimyth/compass-lora-v2","axes":{"intelligence":28.917433770533798,"calibration":83.26568672143415,"speed":83.0508382606611,"cost":72.14079174579808},"capability":{"score":56.09156024598397,"rank":195,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":161,"api_board_rank":null},"composite_score":18.45610384623482,"rank":219,"open_board_rank":176,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008484773925322034,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,282,261 input + 0 output tokens over 1,485 decisions = 863.475420875 tokens/decision. GPU-time cost USD 0.00848477392532203/1,000; old list-price cost USD 0.0259042626262626/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00848477392532203/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.2566348456311971,"p50_s_adjusted":0.6632696912623942,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2566348456311971,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.29841866847127674,"p95_s_adjusted":0.7468373369425535,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.614153784977713,"capability":56.09156024598397,"axes":{"calibration":83.26568672143415,"cost":57.598862966277466,"intelligence":28.917433770533798,"speed":83.28951948101809},"rank":144,"open_board_rank":108,"api_board_rank":null}},{"key":"cribl-decision-1.0-27b","name":"cribl-decision-1.0-27B (Cribl)","board":"open","source_url":"https://huggingface.co/cribl-ai/cribl-decision-1.0-27B","axes":{"intelligence":58.61569357291603,"calibration":89.64583356514017,"speed":86.46160365936547,"cost":64.04033264630576},"capability":{"score":74.13076356902809,"rank":70,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":51,"api_board_rank":null},"composite_score":72.20702429804436,"rank":72,"open_board_rank":55,"api_board_rank":null,"measured_in":"v1.8.3","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.3","measured_on":"2026-10-31","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01579994511966609,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 457,748 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0157999451196661/1,000; old list-price estimate USD 0.0599468502350571/1,000 (qwen/qwen3.5-27b at USD 0.195/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0157999451196661/1,000. Estimated, not a bill. USD 0.0157999451 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0292562886 over all 1,500 items)."},"latency":{"p50_s_raw":0.11137269100004232,"p50_s_adjusted":0.37274538200008467,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.11137269100004232,"latency_basis":"api-end-to-end","p95_s_raw":0.227968423300058,"p95_s_adjusted":0.605936846600116,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"cribl-decision-1.0-9b","name":"cribl-decision-1.0-9B (Cribl)","board":"open","source_url":"https://huggingface.co/cribl-ai/cribl-decision-1.0-9B","axes":{"intelligence":40.91618092554528,"calibration":86.06695892963373,"speed":91.13304559714123,"cost":75.030862863614},"capability":{"score":63.4915699275895,"rank":144,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":113,"api_board_rank":null},"composite_score":44.37738944947674,"rank":155,"open_board_rank":123,"api_board_rank":null,"measured_in":"v1.8.3","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.3","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006796801259584273,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 457,748 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00679680125958427/1,000; old list-price estimate USD 0.0307419744795164/1,000 (qwen/qwen3.5-9b at USD 0.1/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00679680125958427/1,000. Estimated, not a bill. USD 0.00679680126 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0125854348 over all 1,500 items)."},"latency":{"p50_s_raw":0.044512586999985615,"p50_s_adjusted":0.23902517399997122,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.044512586999985615,"latency_basis":"api-end-to-end","p95_s_raw":0.08614684295004052,"p95_s_adjusted":0.322293685900081,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"cribl-decision-2","name":"cribl-decision-2 (Cribl)","board":"open","source_url":"https://huggingface.co/cribl-ai/cribl-decision-2","axes":{"intelligence":38.757150260622026,"calibration":86.72569890176682,"speed":91.18391098931158,"cost":83.11807294295589},"capability":{"score":62.74142458119442,"rank":151,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":119,"api_board_rank":null},"composite_score":39.837192650899226,"rank":174,"open_board_rank":137,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0036536833147096295,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 553,650 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00365368331470963/1,000; old list-price estimate USD 0.0111548018804567/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00365368331470963/1,000. Estimated, not a bill. USD 0.00365368331 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00364340438 over all 1,500 items)."},"latency":{"p50_s_raw":0.04352205550000576,"p50_s_adjusted":0.23704411100001152,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.04352205550000576,"latency_basis":"api-end-to-end","p95_s_raw":0.08560155094997413,"p95_s_adjusted":0.3212031018999483,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"cribl-decision-2-mini","name":"cribl-decision-2-mini (Cribl)","board":"open","source_url":"https://huggingface.co/cribl-ai/cribl-decision-2-mini","axes":{"intelligence":18.61656986771406,"calibration":83.0912038529358,"speed":93.102234850062,"cost":89.44652427171069},"capability":{"score":50.853886860324934,"rank":216,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":180,"api_board_rank":null},"composite_score":6.325019350467616,"rank":258,"open_board_rank":209,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0022479286593751003,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 553,650 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0022479286593751/1,000; old list-price estimate USD 0.00371826729348556/1,000 (deepinfra:Qwen/Qwen3.5-0.8B at USD 0.01/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0022479286593751/1,000. Estimated, not a bill. USD 0.00224792866 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00224160455 over all 1,500 items)."},"latency":{"p50_s_raw":0.029303980999998203,"p50_s_adjusted":0.2086079619999964,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.029303980999998203,"latency_basis":"api-end-to-end","p95_s_raw":0.04233177454995029,"p95_s_adjusted":0.23466354909990056,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"cygnet","name":"Cygnet (blockbrain, frozen Gemma-4-12B-it)","board":"open","source_url":"https://github.com/blockbrain-ai/cygnet-recipe","axes":{"intelligence":54.83928122751086,"calibration":86.96753988280044,"speed":91.37843430750362,"cost":66.80610984255793},"capability":{"score":70.90341055515565,"rank":95,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":72,"api_board_rank":null},"composite_score":71.88318248005169,"rank":79,"open_board_rank":62,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01277800357341222,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 920,404 input + 1,624 output tokens over 1,624 decisions = 567.751231527 tokens/decision. GPU-time cost USD 0.0127780035734122/1,000; old list-price cost USD 0.0283375615763547/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0127780035734122/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04474326502531767,"p50_s_adjusted":0.23948653005063533,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04474326502531767,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07700066294055431,"p95_s_adjusted":0.3040013258811086,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.55012424151383,"capability":70.90341055515565,"axes":{"calibration":86.96753988280044,"cost":56.42912569481113,"intelligence":54.83928122751086,"speed":91.77283664982562},"rank":33,"open_board_rank":16,"api_board_rank":null}},{"key":"d1a-gemma4-e2b","name":"D1A on Gemma 4 E2B (JohnP1 d1a-e2b)","board":"open","source_url":"https://huggingface.co/JohnP1/d1a-e2b","axes":{"intelligence":8.855548462175122,"calibration":82.13235390405904,"speed":93.26967906722538,"cost":84.02860926700725},"capability":{"score":45.49395118311708,"rank":232,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":195,"api_board_rank":null},"composite_score":0.8493910917005856,"rank":272,"open_board_rank":220,"api_board_rank":null,"measured_in":"v1.7.7","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.7","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0034070592886537333,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 664,666 input + 172,216 output tokens over 1,485 decisions = 563.556902357 tokens/decision. GPU-time cost USD 0.00340705928865373/1,000; old list-price cost USD 0.0205487676767677/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00340705928865373/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03238873700001932,"p50_s_adjusted":0.21477747400003863,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03238873700001932,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.034651194550065156,"p95_s_adjusted":0.2193023891001303,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.7","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"d1a-gemma4-e4b","name":"D1A on Gemma 4 E4B (JohnP1 d1a-e4b)","board":"open","source_url":"https://huggingface.co/JohnP1/d1a-e4b","axes":{"intelligence":19.912071787053208,"calibration":82.30108318464288,"speed":92.42822411239143,"cost":81.21501802704326},"capability":{"score":51.10657748584804,"rank":215,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":179,"api_board_rank":null},"composite_score":7.4194035339710895,"rank":254,"open_board_rank":206,"api_board_rank":null,"measured_in":"v1.7.7","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.7","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004228302654292637,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 664,666 input + 172,302 output tokens over 1,485 decisions = 563.614814815 tokens/decision. GPU-time cost USD 0.00422830265429264/1,000; old list-price cost USD 0.0205545589225589/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00422830265429264/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04040702399998963,"p50_s_adjusted":0.23081404799997926,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04040702399998963,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0488469613500456,"p95_s_adjusted":0.2476939227000912,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.7","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decide-1.1","name":"decide-1.1 (decide.sh, API)","board":"api","source_url":null,"axes":{"intelligence":72.84846318590664,"calibration":90.42055602589686,"speed":90.54529257210464,"cost":70.21372225584122},"capability":{"score":81.63450960590174,"rank":25,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":7},"composite_score":79.88339520691216,"rank":2,"open_board_rank":null,"api_board_rank":1,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.009837300201477502,"basis":"API launch list price (I-1): https://decide.sh/pricing, applied since 2026-10-29. USD 0.03 per million input tokens; output free (input-only billing), no base model declared and no override. The frozen pricing rules apply this public launch price. Native server usage; 19,926 reported output tokens are unbilled. USD 0.0098373002 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0177799 over all 1,500 items)."},"latency":{"p50_s_raw":0.09421851398656145,"p50_s_adjusted":0.09421851398656145,"adjustment":"none (API)","p50_s_measured":0.09421851398656145,"latency_basis":"api-end-to-end","p95_s_raw":0.8820043819549017,"p95_s_adjusted":0.8820043819549017,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decider-12b","name":"decider-12b v2 (Mapika)","board":"open","source_url":"https://huggingface.co/Mapika/decider-12b","axes":{"intelligence":61.94396690132827,"calibration":80.75417276143466,"speed":90.45532474538248,"cost":68.15540194207952},"capability":{"score":71.34906983138146,"rank":88,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":65,"api_board_rank":null},"composite_score":73.72671792000273,"rank":50,"open_board_rank":36,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01152090517060813,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 831,319 input + 0 output tokens over 1,624 decisions = 511.895935961 tokens/decision. GPU-time cost USD 0.0115209051706081/1,000; old list-price cost USD 0.0255947967980296/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0115209051706081/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05426313355565071,"p50_s_adjusted":0.25852626711130144,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05426313355565071,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09915375960990787,"p95_s_adjusted":0.34830751921981573,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":23,"n_over_limit_answered_ok":23,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.87955164338175,"capability":72.50047627291733,"axes":{"calibration":81.98845142422223,"cost":57.755449421401664,"intelligence":63.01250112161242,"speed":90.47583056481872},"rank":21,"open_board_rank":8,"api_board_rank":null}},{"key":"decider-12b-v1","name":"decider-12b v1, stock Gemma-4-12B-it (Mapika)","board":"open","source_url":"https://huggingface.co/Mapika/decider-12b","axes":{"intelligence":59.571985034161884,"calibration":82.51700722651287,"speed":90.44187377539939,"cost":68.15540194207952},"capability":{"score":71.04449613033738,"rank":94,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":71,"api_board_rank":null},"composite_score":73.21407386805642,"rank":58,"open_board_rank":43,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01152090517060813,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 831,319 input + 0 output tokens over 1,624 decisions = 511.895935961 tokens/decision. GPU-time cost USD 0.0115209051706081/1,000; old list-price cost USD 0.0255947967980296/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0115209051706081/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.054376568645238876,"p50_s_adjusted":0.2587531372904778,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.054376568645238876,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09954081624746322,"p95_s_adjusted":0.3490816324949264,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":23,"n_over_limit_answered_ok":23,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.42352588714175,"capability":72.1959364423342,"axes":{"calibration":83.7629373406296,"cost":57.755449421401664,"intelligence":60.62893554403879,"speed":90.47654290454523},"rank":25,"open_board_rank":10,"api_board_rank":null}},{"key":"decider-2b","name":"decider-2b (Mapika)","board":"open","source_url":"https://huggingface.co/Mapika/decider-2b","axes":{"intelligence":20.654576369710338,"calibration":68.52438190259257,"speed":94.47902882434015,"cost":85.78804188568367},"capability":{"score":44.58947913615145,"rank":234,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":197,"api_board_rank":null},"composite_score":8.006807575798623,"rank":253,"open_board_rank":205,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0029766785445500044,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 799,606 input + 0 output tokens over 1,624 decisions = 492.368226601 tokens/decision. GPU-time cost USD 0.00297667854455/1,000; old list-price cost USD 0.0147710467980296/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00297667854455/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.01582353049889207,"p50_s_adjusted":0.18164706099778413,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.01582353049889207,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.02313834875822067,"p95_s_adjusted":0.19627669751644133,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":7.6704332123718,"capability":44.58947913615145,"axes":{"calibration":68.52438190259257,"cost":64.91766177740826,"intelligence":20.654576369710338,"speed":94.55945724666857},"rank":178,"open_board_rank":139,"api_board_rank":null}},{"key":"decider-35b-a3b","name":"decider-35b-a3b (Mapika)","board":"open","source_url":"https://huggingface.co/Mapika/decider-35b-a3b","axes":{"intelligence":48.72471958402083,"calibration":83.61830413355554,"speed":91.16871958052303,"cost":73.18143549566516},"capability":{"score":66.1715118587882,"rank":127,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":100,"api_board_rank":null},"composite_score":66.50611544932869,"rank":123,"open_board_rank":96,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.007833433305248303,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 799,606 input + 0 output tokens over 1,624 decisions = 492.368226601 tokens/decision. GPU-time cost USD 0.0078334333052483/1,000; old list-price cost USD 0.153865070812808/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0078334333052483/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.051989168045111,"p50_s_adjusted":0.253978336090222,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.051989168045111,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07541854126611724,"p95_s_adjusted":0.3008370825322345,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":24.772887446577812,"capability":66.1715118587882,"axes":{"calibration":83.61830413355554,"cost":34.38579876859529,"intelligence":48.72471958402083,"speed":91.27357492980725},"rank":112,"open_board_rank":79,"api_board_rank":null}},{"key":"decider-4b-v2","name":"decider-4b v2 (Mapika)","board":"open","source_url":"https://github.com/Mapika/decider","axes":{"intelligence":40.12597268917964,"calibration":89.63735637833331,"speed":90.31631787234045,"cost":79.08569750994965},"capability":{"score":64.88166453375648,"rank":133,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":105,"api_board_rank":null},"composite_score":43.08354714846196,"rank":163,"open_board_rank":129,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004979014790001528,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 822,885 input + 0 output tokens over 1,624 decisions = 506.702586207 tokens/decision. GPU-time cost USD 0.00497901479000153/1,000; old list-price cost USD 0.0152010775862069/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00497901479000153/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02826417936012149,"p50_s_adjusted":0.20652835872024297,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02826417936012149,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.15009120358154174,"p95_s_adjusted":0.45018240716308344,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":41.240894350857424,"capability":64.88166453375648,"axes":{"calibration":89.63735637833331,"cost":64.54376873042901,"intelligence":40.12597268917964,"speed":91.78549329698745},"rank":80,"open_board_rank":53,"api_board_rank":null}},{"key":"decider-chat-gemma4-31b","name":"decider chat on Gemma-4-31B-it (Mapika, frozen base, inference technique)","board":"open","source_url":"https://github.com/Mapika/decider","axes":{"intelligence":58.54290392078144,"calibration":82.34718405098182,"speed":85.76268320171138,"cost":56.10708393368468},"capability":{"score":70.44504398588163,"rank":101,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":75,"api_board_rank":null},"composite_score":68.13342031667979,"rank":114,"open_board_rank":90,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029046726949891646,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 717,195 input + 0 output tokens over 1,477 decisions = 485.57549086 tokens/decision. GPU-time cost USD 0.0290467269498916/1,000; old list-price cost USD 0.0437017941773866/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0290467269498916/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12592082849005237,"p50_s_adjusted":0.40184165698010477,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12592082849005237,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2551009150512982,"p95_s_adjusted":0.6602018301025965,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":7,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":66.07956202755186,"capability":70.58267678405554,"axes":{"calibration":82.62244964732962,"cost":50.78502198104071,"intelligence":58.54290392078144,"speed":85.78163968405963},"rank":46,"open_board_rank":25,"api_board_rank":null}},{"key":"decisio-gemma-4-12b-tuned-commit","name":"Decisio 12B tuned commit (variant of the Decisio family)","board":"open","source_url":"https://huggingface.co/tachara-ai/Decisio-12b-tuned-commit","axes":{"intelligence":62.78089467434111,"calibration":88.61282851120822,"speed":90.3121994607612,"cost":66.96709724088544},"capability":{"score":75.69686159277467,"rank":62,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":47,"api_board_rank":null},"composite_score":75.16131305837865,"rank":27,"open_board_rank":17,"api_board_rank":null,"measured_in":"v1.7.6","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.6","measured_on":"2026-10-25","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.012621086947699698,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 831,272 input + 1,485 output tokens over 1,485 decisions = 560.779124579 tokens/decision. GPU-time cost USD 0.0126210869476997/1,000; old list-price cost USD 0.0281389562289562/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0126210869476997/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05839707199999111,"p50_s_adjusted":0.26679414399998225,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05839707199999111,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09941095715000756,"p95_s_adjusted":0.3488219143000151,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.6","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisio-gemma-4-12b-tuned-commit-think","name":"Decisio 12B tuned commit-think (variant of the Decisio family)","board":"open","source_url":"https://huggingface.co/tachara-ai/Decisio-12b-tuned-commit-think","axes":{"intelligence":65.73656738918622,"calibration":86.83008860238174,"speed":90.60515121372691,"cost":70.73054063499694},"capability":{"score":76.28332799578398,"rank":60,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":45,"api_board_rank":null},"composite_score":77.06528795006618,"rank":14,"open_board_rank":9,"api_board_rank":null,"measured_in":"v1.7.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.14","measured_on":"2026-10-27","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009454719427207697,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 616,406 input + 9,110 output tokens over 1,489 decisions = 420.091336467 tokens/decision. GPU-time cost USD 0.0094547194272077/1,000; old list-price cost USD 0.0216163868368032/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0094547194272077/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05185191250001253,"p50_s_adjusted":0.2537038250000251,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05185191250001253,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0964462016999846,"p95_s_adjusted":0.3428924033999692,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.14","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisio-gemma-4-12b-v080","name":"decisio v0.8.0 on gemma-4-12B-it (frozen, one forward pass, self-hosted)","board":"open","source_url":"https://github.com/aminry/decisio","axes":{"intelligence":52.324988528776075,"calibration":89.15794375860679,"speed":86.316096866496,"cost":67.72536476642883},"capability":{"score":70.74146614369144,"rank":97,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":73,"api_board_rank":null},"composite_score":70.5738473018771,"rank":90,"open_board_rank":70,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011907516019509904,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,965 input + 1,477 output tokens over 1,477 decisions = 529.07379824 tokens/decision. GPU-time cost USD 0.0119075160195099/1,000; old list-price cost USD 0.0265536899119838/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0119075160195099/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11572447800426744,"p50_s_adjusted":0.3814489560085349,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11572447800426744,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.23114268517820163,"p95_s_adjusted":0.6122853703564033,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.17346094077045,"capability":70.74146614369144,"axes":{"calibration":89.15794375860679,"cost":59.34583145470655,"intelligence":52.324988528776075,"speed":86.9860428376769},"rank":36,"open_board_rank":17,"api_board_rank":null}},{"key":"decisio-gemma-4-12b-v090","name":"decisio v0.9.0 on gemma-4-12B-it (frozen, one prefill per question, self-hosted)","board":"open","source_url":"https://github.com/aminry/decisio","axes":{"intelligence":52.299106200484864,"calibration":87.60965425617285,"speed":85.2750295056469,"cost":67.72536476642883},"capability":{"score":69.95438022832886,"rank":106,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":80,"api_board_rank":null},"composite_score":70.14180529430709,"rank":97,"open_board_rank":75,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011907516019509904,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,965 input + 1,477 output tokens over 1,477 decisions = 529.07379824 tokens/decision. GPU-time cost USD 0.0119075160195099/1,000; old list-price cost USD 0.0265536899119838/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0119075160195099/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.14369468999211676,"p50_s_adjusted":0.43738937998423355,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.14369468999211676,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2643116880877642,"p95_s_adjusted":0.6786233761755284,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.76072338665118,"capability":69.95438022832886,"axes":{"calibration":87.60965425617285,"cost":59.34583145470655,"intelligence":52.299106200484864,"speed":85.86778537116794},"rank":39,"open_board_rank":19,"api_board_rank":null}},{"key":"decisio-gemma-4-31b-v080","name":"decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted)","board":"open","source_url":"https://github.com/aminry/decisio","axes":{"intelligence":70.42321091828492,"calibration":88.71627256952715,"speed":91.13376683561437,"cost":54.98929904529576},"capability":{"score":79.56974174390604,"rank":40,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":30,"api_board_rank":null},"composite_score":73.21985514490987,"rank":57,"open_board_rank":42,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03164875996232158,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,965 input + 1,477 output tokens over 1,477 decisions = 529.07379824 tokens/decision. GPU-time cost USD 0.0316487599623216/1,000; old list-price cost USD 0.0478666418415707/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0316487599623216/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04704356682486832,"p50_s_adjusted":0.24408713364973664,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04704356682486832,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.082778721768409,"p95_s_adjusted":0.315557443536818,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":71.6926576652985,"capability":79.56974174390604,"axes":{"calibration":88.71627256952715,"cost":51.66530928015878,"intelligence":70.42321091828492,"speed":91.18580810090454},"rank":16,"open_board_rank":5,"api_board_rank":null}},{"key":"decisio-qwen3.6-35b-a3b-v090","name":"Decisio v0.9.0 on Qwen3.6-35B-A3B (official FP8, self-hosted)","board":"open","source_url":"https://github.com/aminry/decisio","axes":{"intelligence":33.9677591411048,"calibration":89.00828628793406,"speed":90.99460763280192,"cost":61.71356220552948},"capability":{"score":61.48802271451943,"rank":163,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":131,"api_board_rank":null},"composite_score":27.201182936525417,"rank":198,"open_board_rank":157,"api_board_rank":null,"measured_in":"v1.7.10","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.10","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01888924471985142,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,761,623 input + 1,485 output tokens over 1,485 decisions = 1187.27811448 tokens/decision. GPU-time cost USD 0.0188892447198514/1,000; old list-price cost USD 0.178941717171717/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0188892447198514/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06378673799997614,"p50_s_adjusted":0.2775734759999523,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06378673799997614,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0682620881500725,"p95_s_adjusted":0.286524176300145,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.10","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decision-2-0-vega-27b","name":"Decision 2.0 Vega 27B (vLLM-SR, Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/vllm-sr/Decision-2.0-Vega-27B","axes":{"intelligence":60.4093162408047,"calibration":82.20136692745442,"speed":83.34942176442772,"cost":57.00207511798864},"capability":{"score":71.30534158412956,"rank":89,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":66,"api_board_rank":null},"composite_score":68.6578700666544,"rank":109,"open_board_rank":85,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.027118406268461865,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,329 input + 0 output tokens over 1,477 decisions = 527.643195667 tokens/decision. GPU-time cost USD 0.0271184062684619/1,000; old list-price cost USD 0.221610142180095/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0271184062684619/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1961267376318574,"p50_s_adjusted":0.5422534752637148,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1961267376318574,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3514081372413784,"p95_s_adjusted":0.8528162744827569,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.883756282485887,"capability":71.30534158412956,"axes":{"calibration":82.20136692745442,"cost":29.632311028770005,"intelligence":60.4093162408047,"speed":83.80610555916236},"rank":136,"open_board_rank":101,"api_board_rank":null}},{"key":"decision-2b","name":"Decision 2B (FlyMy.AI, v59)","board":"open","source_url":"https://huggingface.co/flymy-ai/decision-2b-preview","axes":{"intelligence":22.875473256337628,"calibration":86.77033341356835,"speed":90.39467608500453,"cost":66.37275756963365},"capability":{"score":54.82290333495299,"rank":202,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":168,"api_board_rank":null},"composite_score":10.289709237720572,"rank":246,"open_board_rank":199,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.013210160098522166,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.07500718400115147,"p50_s_adjusted":0.30001436800230297,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07500718400115147,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07718096784083173,"p95_s_adjusted":0.3043619356816635,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":10.289708493744515,"capability":54.82290333495299,"axes":{"calibration":86.77033341356835,"cost":66.37275756963365,"intelligence":22.875473256337628,"speed":90.39462801231525},"rank":170,"open_board_rank":131,"api_board_rank":null}},{"key":"decision-4b-v11","name":"Decision 4B v1.1 (FlyMyJev, Qwen3.5-4B + LoRA)","board":"open","source_url":"https://benchmarkheaven.com/jev-models/decision-4b-v11","axes":{"intelligence":29.134168280172865,"calibration":85.23934158755043,"speed":93.91870342645993,"cost":77.61552273233744},"capability":{"score":57.18675493386165,"rank":182,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":148,"api_board_rank":null},"composite_score":19.516239094521907,"rank":216,"open_board_rank":173,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.019326324458234012,"p50_s_adjusted":0.18865264891646802,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.019326324458234012,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.032507002749480304,"p95_s_adjusted":0.2150140054989606,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.718356274892123,"capability":57.18675493386165,"axes":{"calibration":85.23934158755043,"cost":63.07359395281681,"intelligence":29.134168280172865,"speed":93.95641941218838},"rank":139,"open_board_rank":104,"api_board_rank":null}},{"key":"decision-4b-v12","name":"Decision 4B v1.2 (FlyMyJev, Qwen3.5-4B + LoRA)","board":"open","source_url":"https://benchmarkheaven.com/jev-models/decision-4b-v12","axes":{"intelligence":32.40310972936693,"calibration":88.42488519968094,"speed":93.8753186824915,"cost":77.61552273233744},"capability":{"score":60.41399746452394,"rank":168,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":136,"api_board_rank":null},"composite_score":25.567225216905292,"rank":203,"open_board_rank":162,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.01960996154230088,"p50_s_adjusted":0.18921992308460175,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.01960996154230088,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0332608105847612,"p95_s_adjusted":0.2165216211695224,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":24.463556699266523,"capability":60.41399746452394,"axes":{"calibration":88.42488519968094,"cost":63.07359395281681,"intelligence":32.40310972936693,"speed":93.92936299294371},"rank":115,"open_board_rank":82,"api_board_rank":null}},{"key":"decision-fast","name":"Decision Fast (FlyMy.AI, v53a)","board":"open","source_url":"https://huggingface.co/flymy-ai/decision-fast-preview","axes":{"intelligence":6.110691078442272,"calibration":75.07737238688318,"speed":91.47223870429099,"cost":86.6432611441359},"capability":{"score":40.59403173266273,"rank":250,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":213,"api_board_rank":null},"composite_score":0.29956154664320095,"rank":283,"open_board_rank":230,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0027875625965184925,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.6B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,805 input + 0 output tokens over 1,624 decisions = 461.08682266 tokens/decision. GPU-time cost USD 0.00278756259651849/1,000; old list-price cost USD 0.00461086822660099/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00278756259651849/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05762279196642339,"p50_s_adjusted":0.2652455839328468,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05762279196642339,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05930679076118395,"p95_s_adjusted":0.2686135815223679,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.2981489585308357,"capability":40.59403173266273,"axes":{"calibration":75.07737238688318,"cost":80.08651867745036,"intelligence":6.110691078442272,"speed":91.472185793536},"rank":210,"open_board_rank":166,"api_board_rank":null}},{"key":"decision-king-v1.1","name":"Decision-King v1.1 (xsdfasfgsa)","board":"open","source_url":"https://huggingface.co/xsdfasfgsa/Decision-King-v1.1","axes":{"intelligence":72.81928758511125,"calibration":86.93485761485668,"speed":82.77728277705924,"cost":61.07389552953992},"capability":{"score":79.87707259998396,"rank":36,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":27,"api_board_rank":null},"composite_score":74.50217464703219,"rank":36,"open_board_rank":24,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.019839778339703577,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 574,788 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0198397783397036/1,000; old list-price estimate USD 0.16212959032908/1,000 (qwen/qwen3.8-27b at USD 0.42/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0198397783397036/1,000. Estimated, not a bill. USD 0.0198397783 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0333017193 over all 1,500 items)."},"latency":{"p50_s_raw":0.22799053500000355,"p50_s_adjusted":0.6059810700000071,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.22799053500000355,"latency_basis":"api-end-to-end","p95_s_raw":0.3602939879999724,"p95_s_adjusted":0.8705879759999449,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decision-machine-1","name":"decision-machine-1 (milliseconds.ai)","board":"api","source_url":"https://www.milliseconds.ai","axes":{"intelligence":13.11,"calibration":79.15,"speed":92.87,"cost":58.33},"capability":{"score":46.13,"rank":230,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":36},"composite_score":2.35,"rank":268,"open_board_rank":null,"api_board_rank":51,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.02449299492385787,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.04/M input, 0/M output (spec decision-machine-1.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.171,"p50_s_adjusted":0.171,"adjustment":"none (API)","p50_s_measured":0.171,"latency_basis":"api-end-to-end","p95_s_raw":0.302,"p95_s_adjusted":0.302,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":2.35,"capability":46.13,"axes":{"calibration":79.15,"cost":58.33,"intelligence":13.11,"speed":92.97},"rank":196,"open_board_rank":null,"api_board_rank":42}},{"key":"decision2-lux-9b","name":"Decision 2.0 Lux 9B (vLLM-SR, Qwen3.5-9B lineage)","board":"open","source_url":"https://huggingface.co/vllm-sr/Decision-2.0-Lux-9B","axes":{"intelligence":35.82270435669186,"calibration":87.47631981746561,"speed":89.53277608263193,"cost":68.1128992661464},"capability":{"score":61.64951208707873,"rank":161,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":129,"api_board_rank":null},"composite_score":31.492390122627054,"rank":186,"open_board_rank":147,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011558550047406203,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 776,349 input + 0 output tokens over 1,485 decisions = 522.793939394 tokens/decision. GPU-time cost USD 0.0115585500474062/1,000; old list-price cost USD 0.0522793939393939/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0115585500474062/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07030085151200183,"p50_s_adjusted":0.2906017030240037,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07030085151200183,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11659944246493978,"p95_s_adjusted":0.3831988849298795,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":27.113058507774674,"capability":61.64951208707873,"axes":{"calibration":87.47631981746561,"cost":48.45008367155653,"intelligence":35.82270435669186,"speed":89.93254869015362},"rank":108,"open_board_rank":75,"api_board_rank":null}},{"key":"decisionnode-edge-0.9.2","name":"DecisionNode Edge 0.9.2","board":"open","source_url":"https://huggingface.co/bynn-intelligence-com/decisionnode-edge","axes":{"intelligence":26.810082223231646,"calibration":77.47576560172445,"speed":92.88928290241705,"cost":86.77310580145102},"capability":{"score":52.14292391247805,"rank":213,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":178,"api_board_rank":null},"composite_score":15.863513503818993,"rank":233,"open_board_rank":187,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0027599198706643154,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 677,924 input + 0 output tokens over 1,485 decisions = 456.514478114 tokens/decision. GPU-time cost USD 0.00275991987066432/1,000; old list-price cost USD 0.00913028956228956/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00275991987066432/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.031239357499998732,"p50_s_adjusted":0.21247871499999746,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.031239357499998732,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04598353859999804,"p95_s_adjusted":0.24196707719999608,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisionnode-edge-mini-0.9.2","name":"DecisionNode Edge Mini 0.9.2","board":"open","source_url":"https://huggingface.co/bynn-intelligence-com/decisionnode-edge-mini","axes":{"intelligence":17.55610652928715,"calibration":77.66777221653875,"speed":92.90461510807665,"cost":86.77310580145102},"capability":{"score":47.61193937291295,"rank":227,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":191,"api_board_rank":null},"composite_score":5.353100170436692,"rank":259,"open_board_rank":210,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0027599198706643154,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 677,924 input + 0 output tokens over 1,485 decisions = 456.514478114 tokens/decision. GPU-time cost USD 0.00275991987066432/1,000; old list-price cost USD 0.00456514478114478/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00275991987066432/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03138889749999407,"p50_s_adjusted":0.21277779499998814,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03138889749999407,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04538772000000506,"p95_s_adjusted":0.2407754400000101,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisionnode-flash-0.9.1-rc1","name":"DecisionNode Flash","board":"api","source_url":"https://decisionnode.com","axes":{"intelligence":45.17690085075341,"calibration":86.74280155989597,"speed":90.91186417999961,"cost":null},"capability":{"score":65.9598512053247,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.7.19","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.19","measured_on":"2026-10-28","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Not ranked because it has no price: under the frozen pricing rules a hosted model's own tariff counts only once it is admitted to the frozen list-price set, and this model declares no base model, so it has no cost axis and no composite.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"NOT PRICED. The vendor states USD 0.042 (DecisionNode) / 0.035 (Flash) per 1M input tokens, output not billed, but under the frozen pricing rules a hosted model's own tariff counts only once it is admitted to the frozen list-price set, and these models declare no base model. Token usage on the fixed 1,489-item cost set: 471,195 input, 0 output (473,406 input on all 1,495 answered items). The vendor tariff is not used to compute a cost."},"latency":{"p50_s_raw":0.256085725966841,"p50_s_adjusted":0.256085725966841,"adjustment":"none (API)","p50_s_measured":0.256085725966841,"latency_basis":"api-end-to-end","p95_s_raw":0.31653971727064345,"p95_s_adjusted":0.31653971727064345,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.19","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisionnode-main-0.9.1-rc1","name":"DecisionNode","board":"api","source_url":"https://decisionnode.com","axes":{"intelligence":58.166046525064246,"calibration":90.62294768982271,"speed":82.66643460386248,"cost":null},"capability":{"score":74.39449710744347,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.7.19","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.19","measured_on":"2026-10-28","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Not ranked because it has no price: under the frozen pricing rules a hosted model's own tariff counts only once it is admitted to the frozen list-price set, and this model declares no base model, so it has no cost axis and no composite.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"NOT PRICED. The vendor states USD 0.042 (DecisionNode) / 0.035 (Flash) per 1M input tokens, output not billed, but under the frozen pricing rules a hosted model's own tariff counts only once it is admitted to the frozen list-price set, and these models declare no base model. Token usage on the fixed 1,489-item cost set: 471,195 input, 0 output (473,406 input on all 1,495 answered items). The vendor tariff is not used to compute a cost."},"latency":{"p50_s_raw":0.6751686710049398,"p50_s_adjusted":0.6751686710049398,"adjustment":"none (API)","p50_s_measured":0.6751686710049398,"latency_basis":"api-end-to-end","p95_s_raw":0.8015751732105854,"p95_s_adjusted":0.8015751732105854,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.19","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"decisor_4b","name":"decisor-4b","board":"open","source_url":"https://huggingface.co/underlabs/decisor-4b","axes":{"intelligence":39.60641807869822,"calibration":51.86891657577944,"speed":91.86196998743578,"cost":78.6876666032308},"capability":{"score":45.737667327238825,"rank":231,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":194,"api_board_rank":null},"composite_score":36.84364855230138,"rank":181,"open_board_rank":143,"api_board_rank":null,"measured_in":"v1.6.3","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-09","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005133471003269886,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 771,182 input + 1,479 output tokens over 1,479 decisions = 522.421230561 tokens/decision. GPU-time cost USD 0.00513347100326989/1,000; old list-price cost USD 0.0157926369168357/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00513347100326989/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.048687305476050824,"p50_s_adjusted":0.24737461095210164,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.048687305476050824,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.056649096039473074,"p95_s_adjusted":0.2632981920789461,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":35.341565591640844,"capability":45.737667327238825,"axes":{"calibration":51.86891657577944,"cost":64.04636047691767,"intelligence":39.60641807869822,"speed":91.94556701350172},"rank":92,"open_board_rank":62,"api_board_rank":null}},{"key":"deck-4b-v1-0","name":"deck-4B v1.0 (krishna765, JevK5 + LoRA, FP8 weight-only)","board":"open","source_url":"https://github.com/krishna-gogineni-765/deck","axes":{"intelligence":38.62465958342338,"calibration":88.59116686443139,"speed":89.8569997500675,"cost":78.00338436233434},"capability":{"score":63.60791322392738,"rank":143,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":112,"api_board_rank":null},"composite_score":39.04972547531234,"rank":178,"open_board_rank":140,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005410289709263655,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 813,225 input + 0 output tokens over 1,477 decisions = 550.592417062 tokens/decision. GPU-time cost USD 0.00541028970926366/1,000; old list-price cost USD 0.0165177725118483/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00541028970926366/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07722867885604501,"p50_s_adjusted":0.30445735771209004,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07722867885604501,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09472411600872874,"p95_s_adjusted":0.33944823201745744,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":37.26498528662829,"capability":63.60791322392738,"axes":{"calibration":88.59116686443139,"cost":63.46145558281372,"intelligence":38.62465958342338,"speed":89.93848373337222},"rank":88,"open_board_rank":58,"api_board_rank":null}},{"key":"deck31b","name":"deck-31B (krishna765, frozen Gemma-4-31B-it, TorchAO FP8 dynamic)","board":"open","source_url":"https://github.com/krishna-gogineni-765/deck31b","axes":{"intelligence":73.02512554551848,"calibration":82.15636027953371,"speed":86.37611187702757,"cost":55.01396479280669},"capability":{"score":77.5907429125261,"rank":49,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":36,"api_board_rank":null},"composite_score":71.91617991839077,"rank":78,"open_board_rank":61,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03158890028338916,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,964 input + 0 output tokens over 1,477 decisions = 528.073121192 tokens/decision. GPU-time cost USD 0.0315889002833892/1,000; old list-price cost USD 0.0475265809072444/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0315889002833892/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13031065138056874,"p50_s_adjusted":0.4106213027611375,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13031065138056874,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.20548997265286742,"p95_s_adjusted":0.5609799453057348,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.68200118286292,"capability":77.5907429125261,"axes":{"calibration":82.15636027953371,"cost":49.69190284016272,"intelligence":73.02512554551848,"speed":86.69970536196755},"rank":31,"open_board_rank":14,"api_board_rank":null}},{"key":"deem-0.8-v1","name":"Deem 0.8B v1","board":"open","source_url":"https://huggingface.co/LibertAIDAI/deem-0.8-v1","axes":{"intelligence":6.531204174677637,"calibration":29.6278299611924,"speed":84.83282651822053,"cost":86.84458976989441},"capability":{"score":18.07951706793502,"rank":277,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":238,"api_board_rank":null},"composite_score":0.3247463549936592,"rank":282,"open_board_rank":229,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0027448187663714914,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 737,323 input + 0 output tokens over 1,624 decisions = 454.016625616 tokens/decision. GPU-time cost USD 0.00274481876637149/1,000; old list-price cost USD 0.00454016625615764/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00274481876637149/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1644154244568199,"p50_s_adjusted":0.4788308489136398,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1644154244568199,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2681667725555598,"p95_s_adjusted":0.6863335451111197,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.32342918922131697,"capability":18.07951706793502,"axes":{"calibration":29.6278299611924,"cost":80.28784730320886,"intelligence":6.531204174677637,"speed":85.44505463939521},"rank":209,"open_board_rank":165,"api_board_rank":null}},{"key":"deepseek-flash","name":"DeepSeek V4.1 Flash (thinking default)","board":"api","source_url":"https://benchmarkheaven.com/jev-models/deepseek-flash","axes":{"intelligence":84.48,"calibration":94.76,"speed":65.99,"cost":19.56},"capability":{"score":89.62,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 14.9× Jev (v1.5 reference)","latency 3.6× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":6.9,"rank":256,"open_board_rank":null,"api_board_rank":49,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.480105076142132,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.15/M input, 0.6/M output (spec deepseek-flash.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":2.205,"p50_s_adjusted":2.205,"adjustment":"none (API)","p50_s_measured":2.205,"latency_basis":"api-end-to-end","p95_s_raw":11.414,"p95_s_adjusted":11.414,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":34,"n_over_limit_answered_ok":31,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":7.09,"capability":97.35,"axes":{"calibration":100,"cost":19.56,"intelligence":94.71,"speed":68.07},"rank":180,"open_board_rank":null,"api_board_rank":40}},{"key":"diffusion-jev","name":"Diffusion Jev (DiffusionGemma 26B-A4B on patched SGLang, 48-step diffusion readout, self-hosted)","board":"open","source_url":"https://huggingface.co/google/diffusiongemma-26B-A4B-it","axes":{"intelligence":54.860121638521335,"calibration":61.31489461738089,"speed":86.28217783727928,"cost":71.2270380297646},"capability":{"score":58.08750812795111,"rank":178,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":146,"api_board_rank":null},"composite_score":66.48171557572483,"rank":124,"open_board_rank":97,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00910120155482596,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 763,351 input + 7,370 output tokens over 1,477 decisions = 521.815165877 tokens/decision. GPU-time cost USD 0.00910120155482596/1,000; old list-price cost USD 0.0480112322274881/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00910120155482596/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13112960592843592,"p50_s_adjusted":0.41225921185687187,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13112960592843592,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.21048405179753896,"p95_s_adjusted":0.5709681035950779,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":59.330428224994314,"capability":58.08750812795111,"axes":{"calibration":61.31489461738089,"cost":49.559714426383636,"intelligence":54.860121638521335,"speed":86.79193858661813},"rank":58,"open_board_rank":34,"api_board_rank":null}},{"key":"dopest-1.0-31b","name":"DOPEST 1.0 31B","board":"open","source_url":"https://huggingface.co/dopest-studio/dopest-1.0-31b","axes":{"intelligence":73.86361411900745,"calibration":90.20395849508765,"speed":81.97358858059908,"cost":52.643215806695544},"capability":{"score":82.03378630704755,"rank":19,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":14,"api_board_rank":null},"composite_score":71.65979612895147,"rank":81,"open_board_rank":64,"api_board_rank":null,"measured_in":"v1.7.4","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.4","measured_on":"2026-10-25","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03789304216462614,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 8752.36566075 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 814,647 input + 0 output tokens over 1,485 decisions = 548.583838384 tokens/decision. GPU-time cost USD 0.0378930421646261/1,000; old list-price cost USD 0.0493725454545455/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0378930421646261/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.23605015400005414,"p50_s_adjusted":0.6221003080001083,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.23605015400005414,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.43521207174994736,"p95_s_adjusted":1.0204241434998946,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.4","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"drex-max","name":"Drex Max (Nace AI)","board":"api","source_url":"https://drex.nace.ai","axes":{"intelligence":78.89021516142793,"calibration":89.56795021085651,"speed":89.64251650815261,"cost":55.336763284582396},"capability":{"score":84.22908268614222,"rank":12,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":5},"composite_score":75.37621071625281,"rank":25,"open_board_rank":null,"api_board_rank":10,"measured_in":"v1.6.38","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"listed","usd_per_1000_decisions":0.030815878787878788,"basis":"Launch list price: Drex Max public launch list tariff USD 0.06 per million input tokens, output not billed, https://drex.nace.ai/docs/drex/reference/pricing (saved copy sha256 95d1c429, fetched 21 Oct 2026 20:18 UTC; fetched again 22 Oct 2026 01:12 UTC with identical bytes). Counted as an operator standard launch list price by the frozen pricing module pricing_v15 (sha256 a8a05890ddab2d39d1a1bd543f2b5c83842bdc70aa54766788bc89b7a8a448d7), which now carries a launch-list-price line for this product. Applied to provider-reported usage; not a base-model estimate. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 762,693 input / 262,045 output tokens. All-item usage 1,942,633 input / 264,615 output over 1,500 items."},"latency":{"p50_s_raw":0.2538145645521581,"p50_s_adjusted":0.2538145645521581,"adjustment":"none (API)","p50_s_measured":0.2538145645521581,"latency_basis":"api-end-to-end","p95_s_raw":0.42779116374440485,"p95_s_adjusted":0.42779116374440485,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":75.40515618514833,"capability":84.22908268614222,"axes":{"calibration":89.56795021085651,"cost":55.336763284582396,"intelligence":78.89021516142793,"speed":89.80651018227358},"rank":7,"open_board_rank":null,"api_board_rank":6}},{"key":"eikos-27b","name":"Eikos-27B (caiovicentino1, Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/caiovicentino1/eikos-27b","axes":{"intelligence":64.40270761245412,"calibration":89.97714437568723,"speed":87.88773525505371,"cost":56.05951697168831},"capability":{"score":77.18992599407068,"rank":55,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":41,"api_board_rank":null},"composite_score":71.61070910537894,"rank":82,"open_board_rank":65,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029152967442296302,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.0291529674422963/1,000; old list-price cost USD 0.238236465517241/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0291529674422963/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11535292491316795,"p50_s_adjusted":0.38070584982633593,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11535292491316795,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.1386027937987819,"p95_s_adjusted":0.4272055875975638,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.077889022432696,"capability":77.18992599407068,"axes":{"calibration":89.97714437568723,"cost":28.689752882469676,"intelligence":64.40270761245412,"speed":88.08630815059152},"rank":141,"open_board_rank":106,"api_board_rank":null}},{"key":"elephant-1","name":"baselabs elephant-1 (9B)","board":"open","source_url":"https://huggingface.co/baselabs/elephant-1","axes":{"intelligence":55.83724375498836,"calibration":84.65640909439074,"speed":89.7330689398618,"cost":68.07169708446307},"capability":{"score":70.24682642468954,"rank":102,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":76,"api_board_rank":null},"composite_score":71.99915829082904,"rank":76,"open_board_rank":59,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011595160482360806,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 777,323 input + 1,485 output tokens over 1,485 decisions = 524.44983165 tokens/decision. GPU-time cost USD 0.0115951604823608/1,000; old list-price cost USD 0.0524949831649832/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0115951604823608/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07536707949998345,"p50_s_adjusted":0.3007341589999669,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07536707949998345,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10179922215006058,"p95_s_adjusted":0.35359844430012116,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"evalengine-decision-4b","name":"Decision-4B (Eval Engine / Chromia)","board":"open","source_url":"https://huggingface.co/evalengine/decision-4b-gguf","axes":{"intelligence":37.93489123230218,"calibration":80.25524974800695,"speed":90.61339377257747,"cost":80.41969749290011},"capability":{"score":59.09507049015456,"rank":173,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":141,"api_board_rank":null},"composite_score":36.9629278949271,"rank":180,"open_board_rank":142,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004494452037681966,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 742,801 input + 0 output tokens over 1,624 decisions = 457.389778325 tokens/decision. GPU-time cost USD 0.00449445203768197/1,000; old list-price cost USD 0.0137216933497537/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00449445203768197/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05573910451494157,"p50_s_adjusted":0.26147820902988317,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05573910451494157,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09103327156044543,"p95_s_adjusted":0.33206654312089084,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":35.43243109909669,"capability":59.09507049015456,"axes":{"calibration":80.25524974800695,"cost":65.87776871337948,"intelligence":37.93489123230218,"speed":91.06043780349009},"rank":91,"open_board_rank":61,"api_board_rank":null}},{"key":"exaone-4.0-1.2b-jev-v0.3","name":"EXAONE-4.0-1.2B-JEV v0.3 (unofficial fine-tune of LGAI-EXAONE/EXAONE-4.0-1.2B, native option-label softmax, two reads averaged, self-hosted)","board":"open","source_url":"https://huggingface.co/carrtesy/EXAONE-4.0-1.2B-JEV","axes":{"intelligence":14.831288392065307,"calibration":80.52861474872404,"speed":93.81713750966469,"cost":57.919479713067595},"capability":{"score":47.67995157039467,"rank":225,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":189,"api_board_rank":null},"composite_score":3.265807063300941,"rank":265,"open_board_rank":216,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.025274583615436694,"basis":"ESTIMATE (size-class reference; self-hosted open weights, no bookable listing for this base). THE RATE IS RUN 48'S, CARRIED UNCHANGED FOR THE SAME BASE: run 48 priced exaone-4.0-1.2b-jev (v0.2, identical base) at USD 0.02 per 1M input and 0 output, as the HIGHER of the two bracketing published rates, chosen in that direction because the rate was fixed after the measurement existed. Keeping it makes v0.3 directly comparable with v0.2 and takes no new decision. The frozen 25 Sep snapshot still contains NO listing for LGAI-EXAONE/EXAONE-4.0-1.2B (pricing_v15.reference() raises: 'base model lacks an explicit market-listing decision'), and still brackets 1.2B between deepinfra:Qwen/Qwen3.5-0.8B at (0.01, 0.05) and deepinfra:Qwen/Qwen3.5-2B at (0.02, 0.10). Estimated, not charged. PRICING REFERENCE NOTE: LGAI-EXAONE/EXAONE-4.0-1.2B must be adopted into the frozen UNLISTED_BASE_MODELS table before either EXAONE row may be published. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone and any other decision is reversible by arithmetic from the raw."},"latency":{"p50_s_raw":0.020833803500863723,"p50_s_adjusted":0.19166760700172744,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.020833803500863723,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03331972099695122,"p95_s_adjusted":0.21663944199390245,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.266776278413127,"capability":47.67995157039467,"axes":{"calibration":80.52861474872404,"cost":57.919479713067595,"intelligence":14.831288392065307,"speed":94.09940251974388},"rank":191,"open_board_rank":151,"api_board_rank":null}},{"key":"fastino-glide","name":"Fastino GLiDE","board":"api","source_url":null,"axes":{"intelligence":66.1,"calibration":87.95,"speed":77.23,"cost":35.86},"capability":{"score":77.025,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 4.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":30.57,"rank":190,"open_board_rank":null,"api_board_rank":40,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.13735482233502538,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.15/M input, 0/M output (spec fastino-glide.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.852,"p50_s_adjusted":0.852,"adjustment":"none (API)","p50_s_measured":0.852,"latency_basis":"api-end-to-end","p95_s_raw":2.218,"p95_s_adjusted":2.218,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":30.59,"capability":77.02,"axes":{"calibration":87.95,"cost":35.86,"intelligence":66.1,"speed":77.48},"rank":99,"open_board_rank":null,"api_board_rank":32}},{"key":"fastino-glide-1.2-open","name":"Fastino GLiDE 1.2 (open weights, self-hosted)","board":"open","source_url":"https://huggingface.co/fastino/GLiDE","axes":{"intelligence":76.38100890445796,"calibration":93.46414212015411,"speed":85.49903308442846,"cost":58.10251223805742},"capability":{"score":84.92257551230603,"rank":8,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":6,"api_board_rank":null},"composite_score":75.9034188361358,"rank":19,"open_board_rank":12,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.024922001934719162,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11861.3382645 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=trunk-equivalent). The row's own scored cost basis: 724,641 input + 0 output tokens over 1,482 decisions = 488.961538462 tokens/decision. GPU-time cost USD 0.0249220019347192/1,000; old list-price cost USD 0.205363846153846/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0249220019347192/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13967531800000188,"p50_s_adjusted":0.4293506360000038,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13967531800000188,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2532876803999898,"p95_s_adjusted":0.6565753607999796,"n":1497},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1497,"treatment":"share"},"previous":null},{"key":"fastino-glide-api","name":"Fastino GLiDE (hosted API, 22 Oct 2026)","board":"api","source_url":"https://fastino.ai","axes":{"intelligence":76.06842865080436,"calibration":92.84353248115252,"speed":87.00413908296701,"cost":59.111125613009825},"capability":{"score":84.45598056597845,"rank":10,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":4},"composite_score":76.44111138489434,"rank":17,"open_board_rank":null,"api_board_rank":6,"measured_in":"v1.6.40","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"mixed","usd_per_1000_decisions":0.023065479082321188,"basis":"Operator's own published standard usage tariff: USD 0.04 per 1M INPUT tokens, output not billed (docs.fastino.ai/pricing, GLiDE row; page read 22 Oct 2026 20:34 UTC, saved copy sha256 22c0758f). Counted under interpretation I-1 (a manufacturer's standard, non-promotional list price counts from day 1) by the frozen pricing module pricing_v15 (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7), which carries one added launch-list-price key for this row and nothing else. Applied to this row's own provider-reported input tokens on the answered rows; the 17 requests that returned an error are priced at the API's 65,536-token prompt ceiling, the documented context estimate. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common items of this draw, leaving out the 15 very long items that count for no row; this row scored 1,497 items, so the basis here is 1,482 items. This row: USD 0.023065 per 1,000 decisions (0.357x the official cost cap USD 0.064595), against USD 0.049101 over all 1,497 scored items. The older measurement of this vendor's GLiDE on 6 Oct 2026 (key fastino-glide) was priced at the then tariff of USD 0.15 per 1M input tokens and is unchanged."},"latency":{"p50_s_raw":0.35695503698661923,"p50_s_adjusted":0.35695503698661923,"adjustment":"none (API)","p50_s_measured":0.35695503698661923,"latency_basis":"api-end-to-end","p95_s_raw":0.5584349319804464,"p95_s_adjusted":0.5584349319804464,"n":1497},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1497,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":76.60902426459442,"capability":84.7015252595811,"axes":{"calibration":93.01106628088438,"cost":59.111125613009825,"intelligence":76.39198423827783,"speed":87.30503336667806},"rank":2,"open_board_rank":null,"api_board_rank":2}},{"key":"fastino-gliner-2-5-decide","name":"Fastino GLiNER-2.5-Decide (hosted API)","board":"api","source_url":"https://fastino.ai/","axes":{"intelligence":7.011351364482595,"calibration":62.13288114639033,"speed":84.98447762423095,"cost":45.81558630529788},"capability":{"score":34.57211625543646,"rank":261,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":37},"composite_score":0.3433992147410167,"rank":281,"open_board_rank":null,"api_board_rank":53,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"mixed","usd_per_1000_decisions":0.06399515910629654,"basis":"ESTIMATE: published Fastino sources conflict (USD 0.03/M vs 0.15/M input); conservative higher USD 0.15/M input, 0/M output x measured tokens; no invoice"},"latency":{"p50_s_raw":0.42045173048973083,"p50_s_adjusted":0.42045173048973083,"adjustment":"none (API)","p50_s_measured":0.42045173048973083,"latency_basis":"api-end-to-end","p95_s_raw":0.7548073027282952,"p95_s_adjusted":0.7548073027282952,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.34332996901846624,"capability":34.57211625543646,"axes":{"calibration":62.13288114639033,"cost":45.81558630529788,"intelligence":7.011351364482595,"speed":84.7052606179839},"rank":208,"open_board_rank":null,"api_board_rank":44}},{"key":"gemini-3.1-flash-lite","name":"Gemini 3.1 Flash-Lite","board":"api","source_url":"https://benchmarkheaven.com/jev-models/gemini-3.1-flash-lite","axes":{"intelligence":58.57,"calibration":60.93,"speed":78.98,"cost":29.17},"capability":{"score":59.75,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 7.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":16.93,"rank":224,"open_board_rank":null,"api_board_rank":45,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.22953722504230117,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.25/M input, 1.5/M output (spec gemini-3.1-flash-lite.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.962,"p50_s_adjusted":0.962,"adjustment":"none (API)","p50_s_measured":0.962,"latency_basis":"api-end-to-end","p95_s_raw":1.317,"p95_s_adjusted":1.317,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.94,"capability":59.75,"axes":{"calibration":60.93,"cost":29.17,"intelligence":58.57,"speed":79.03},"rank":148,"open_board_rank":null,"api_board_rank":37}},{"key":"gemini-decision-1","name":"Gemini Decision 1","board":"api","source_url":null,"axes":{"intelligence":67.4004381377088,"calibration":89.6397534987547,"speed":90.06659217686632,"cost":52.5184149091935},"capability":{"score":78.52009581823175,"rank":47,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":13},"composite_score":71.25514187539062,"rank":86,"open_board_rank":null,"api_board_rank":19,"measured_in":"v1.8.10","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.10","measured_on":"2026-11-11","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.03825775688381464,"basis":"Google list price USD 0.10 per 1M input tokens, output free (Gemini API paid tier; free tier, batch and caching not counted), times API-reported usage; launch list price counted from day 1; frozen pricing 1b3c52fb, page re-fetched 12 Nov 2026 15:06Z. USD 0.0382578 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0647295 over all 1,500 items); the 3 failed requests count as zero-token items."},"latency":{"p50_s_raw":0.23401184054091573,"p50_s_adjusted":0.23401184054091573,"adjustment":"none (API)","p50_s_measured":0.23401184054091573,"latency_basis":"api-end-to-end","p95_s_raw":0.42082637383136895,"p95_s_adjusted":0.42082637383136895,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.10","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"gev-26b-decide","name":"GEV-26B-Decide (AutoTrust, Gemma-4-26B-A4B + LoRA + head)","board":"open","source_url":"https://huggingface.co/autotrust/GEV-26B-Decide","axes":{"intelligence":49.817428608862244,"calibration":90.33483359631919,"speed":89.61965681094298,"cost":72.5645152590109},"capability":{"score":70.07613110259072,"rank":104,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":78,"api_board_rank":null},"composite_score":70.80379734355978,"rank":89,"open_board_rank":69,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008213271099431558,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 694,051 input + 1,477 output tokens over 1,477 decisions = 470.905890318 tokens/decision. GPU-time cost USD 0.00821327109943156/1,000; old list-price cost USD 0.0425915301286391/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00821327109943156/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.054441039988887496,"p50_s_adjusted":0.258882079977775,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.054441039988887496,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.13581540058075917,"p95_s_adjusted":0.4216308011615183,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":64.2201412904201,"capability":70.07613110259072,"axes":{"calibration":90.33483359631919,"cost":51.12030271996105,"intelligence":49.817428608862244,"speed":89.87310530072351},"rank":50,"open_board_rank":29,"api_board_rank":null}},{"key":"goddragon-arbiter-4b","name":"Arbiter-4B (Goddragon, NeoHorse 4B pointer model, author server, self-hosted)","board":"open","source_url":"https://huggingface.co/Goddragon/Arbiter-4B","axes":{"intelligence":27.391250053452495,"calibration":77.19519806916665,"speed":92.05594930781893,"cost":80.74092951874186},"capability":{"score":52.29322406130957,"rank":212,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":177,"api_board_rank":null},"composite_score":16.509995434698638,"rank":226,"open_board_rank":181,"api_board_rank":null,"measured_in":"v1.6.36","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004384994127077292,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 662,682 input + 0 output tokens over 1,485 decisions = 446.250505051 tokens/decision. GPU-time cost USD 0.00438499412707729/1,000; old list-price cost USD 0.0133875151515152/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00438499412707729/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04713439749991721,"p50_s_adjusted":0.24426879499983442,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04713439749991721,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05249908954998545,"p95_s_adjusted":0.2549981790999709,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.917217056211483,"capability":52.29322406130957,"axes":{"calibration":77.19519806916665,"cost":66.19900073922123,"intelligence":27.391250053452495,"speed":92.16491010646149},"rank":155,"open_board_rank":117,"api_board_rank":null}},{"key":"gpt-5.6-luna","name":"GPT-5.6 Luna (low reasoning effort)","board":"api","source_url":"https://benchmarkheaven.com/jev-models/gpt-5.6-luna","axes":{"intelligence":92.97,"calibration":94.31,"speed":73.21,"cost":30.16},"capability":{"score":93.64,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 6.6× Jev (v1.5 reference)","latency 2.6× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":21.36,"rank":210,"open_board_rank":null,"api_board_rank":42,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.2127451776649746,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.2/M input, 1.2/M output (spec gpt-5.6-luna.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":1.586,"p50_s_adjusted":1.586,"adjustment":"none (API)","p50_s_measured":1.586,"latency_basis":"api-end-to-end","p95_s_raw":3.011,"p95_s_adjusted":3.011,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":4,"n_over_limit_answered_ok":4,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.41,"capability":94.49,"axes":{"calibration":94.91,"cost":30.16,"intelligence":94.07,"speed":73.12},"rank":121,"open_board_rank":null,"api_board_rank":34}},{"key":"gpt-6-luna","name":"GPT-6 Luna (default medium reasoning effort)","board":"api","source_url":"https://benchmarkheaven.com/jev-models/gpt-6-luna","axes":{"intelligence":96.88,"calibration":99.92,"speed":71.38,"cost":38.41},"capability":{"score":98.4,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 3.5× Jev (v1.5 reference)","latency 3.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":39.1,"rank":177,"open_board_rank":null,"api_board_rank":38,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.1129575296108291,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.1/M input, 0.5/M output (spec gpt-6-luna.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":1.932,"p50_s_adjusted":1.932,"adjustment":"none (API)","p50_s_measured":1.932,"latency_basis":"api-end-to-end","p95_s_raw":3.765,"p95_s_adjusted":3.765,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":39.16,"capability":98.44,"axes":{"calibration":100,"cost":38.41,"intelligence":96.88,"speed":71.77},"rank":84,"open_board_rank":null,"api_board_rank":30}},{"key":"gpt-6-luna-low","name":"GPT-6 Luna (low reasoning effort)","board":"api","source_url":"https://benchmarkheaven.com/jev-models/gpt-6-luna-low","axes":{"intelligence":95.66,"calibration":98.48,"speed":71.82,"cost":38.98},"capability":{"score":97.07,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 3.3× Jev (v1.5 reference)","latency 3.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":40.39,"rank":170,"open_board_rank":null,"api_board_rank":36,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.10814365482233503,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.1/M input, 0.5/M output (spec gpt-6-luna-low.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":1.896,"p50_s_adjusted":1.896,"adjustment":"none (API)","p50_s_measured":1.896,"latency_basis":"api-end-to-end","p95_s_raw":3.465,"p95_s_adjusted":3.465,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":3,"n_over_limit_answered_ok":3,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":40.48,"capability":97.82,"axes":{"calibration":99.02,"cost":38.98,"intelligence":96.62,"speed":71.66},"rank":81,"open_board_rank":null,"api_board_rank":28}},{"key":"granite-switch-4.1-30b","name":"Granite Switch 4.1 30B (revision 4.1.1, decision adapter)","board":"open","source_url":"https://huggingface.co/ibm-granite/granite-switch-4.1-30b","axes":{"intelligence":44.59437044734082,"calibration":79.58429831913075,"speed":88.34207613347354,"cost":46.610086510235},"capability":{"score":62.08933438323579,"rank":154,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":122,"api_board_rank":null},"composite_score":40.80390170483756,"rank":168,"open_board_rank":133,"api_board_rank":null,"measured_in":"v1.8.5","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.5","measured_on":"2026-10-31","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.06020932840832774,"basis":"Sibling estimate (*est.): the declared measured base is ibm-granite/granite-4.1-30b; granite-4.2-30b is the sibling reference at USD 0.16 per million input tokens and USD 0.65 per million output tokens. This is not a listing or tariff for the measured base and not an author API quote. The frozen price catalog has no exact 4.1-30b listing. Native server usage includes every full-prompt chunk and every generated label-read token. On the frozen 1,489-decision common set: 554,274 input + 1,489 output tokens, yielding USD 0.0602093284083277 per 1,000 decisions. Across all 1,500 decisions: 818,484 input + 1,500 output tokens, yielding USD 0.08795496 per 1,000. Granite is outside the six calibrated Gemma/Qwen throughput classes; under the existing v1.8 method the sibling estimate is retained and the GPU-time component is unavailable. No cross-family throughput mapping, observed-loop extrapolation or pod bill is used. Estimated, not a bill. USD 0.0602093284 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.08795496 over all 1,500 items)."},"latency":{"p50_s_raw":0.07952452649999486,"p50_s_adjusted":0.30904905299998975,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.07952452649999486,"latency_basis":"api-end-to-end","p95_s_raw":0.16199270275009212,"p95_s_adjusted":0.4739854055001842,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.5","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"granite-switch-4.1-8b","name":"Granite Switch 4.1 8B (revision 4.1.1, decision adapter)","board":"open","source_url":"https://huggingface.co/ibm-granite/granite-switch-4.1-8b","axes":{"intelligence":29.263263741404746,"calibration":80.2886386150226,"speed":92.68935191738173,"cost":59.3855424393103},"capability":{"score":54.775951178213674,"rank":203,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":169,"api_board_rank":null},"composite_score":18.451735459722542,"rank":220,"open_board_rank":177,"api_board_rank":null,"measured_in":"v1.8.4","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.4","measured_on":"2026-10-31","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0225847481531229,"basis":"Sibling estimate (*est.): the declared measured base is ibm-granite/granite-4.1-8b; granite-4.2-8b is the sibling reference at USD 0.06 per million input tokens and USD 0.25 per million output tokens. This is not a listing or tariff for the measured base and not an author API quote. The frozen price catalog has no exact 4.1-8b listing. Native server usage includes every full-prompt chunk and every generated label-read token. On the frozen 1,489-decision common set: 554,274 input + 1,489 output tokens, yielding USD 0.0225847481531229 per 1,000 decisions. Across all 1,500 decisions: 818,484 input + 1,500 output tokens, yielding USD 0.03298936 per 1,000. Granite is outside the six calibrated Gemma/Qwen throughput classes; under the existing v1.8 method the sibling estimate is retained and the GPU-time component is unavailable. No cross-family throughput mapping, observed-loop extrapolation or pod bill is used. Estimated, not a bill. USD 0.0225847482 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.03298936 over all 1,500 items)."},"latency":{"p50_s_raw":0.028615256000023237,"p50_s_adjusted":0.20723051200004647,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.028615256000023237,"latency_basis":"api-end-to-end","p95_s_raw":0.0548916138000095,"p95_s_adjusted":0.259783227600019,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.4","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"gte-reranker-modernbert-base","name":"Alibaba GTE Reranker ModernBERT-base","board":"open","source_url":"https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base","axes":{"intelligence":0,"calibration":83.4556233651092,"speed":93.50537133277697,"cost":69.26965431492843},"capability":{"score":41.7278116825546,"rank":247,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":210,"api_board_rank":null},"composite_score":0,"rank":312,"open_board_rank":259,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.010576570197044334,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: size-class proxy (deepinfra:thenlper/gte-base); no exact base-model floor applies"},"latency":{"p50_s_raw":0.024703769711777568,"p50_s_adjusted":0.19940753942355513,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.024703769711777568,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03686424517072736,"p95_s_adjusted":0.2237284903414547,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0,"capability":41.7278116825546,"axes":{"calibration":83.4556233651092,"cost":69.26965431492843,"intelligence":0,"speed":93.76120791185616},"rank":239,"open_board_rank":195,"api_board_rank":null}},{"key":"gutsy-0.8b-v0.3","name":"Gutsy 0.8B v0.3","board":"open","source_url":"https://huggingface.co/kouhxp/gutsy","axes":{"intelligence":8.600978834309654,"calibration":75.89600269702092,"speed":71.56789504904872,"cost":84.91099734772837},"capability":{"score":42.24849076566529,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.7626893562712093,"rank":274,"open_board_rank":222,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0031839537830687836,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 789,980 input + 0 output tokens over 1,500 decisions = 526.653333333 tokens/decision. GPU-time cost USD 0.00318395378306878/1,000; old list-price cost USD 0.00526653333333333/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00318395378306878/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.875635983499933,"p50_s_adjusted":1.901271966999866,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.875635983499933,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.7578893787498597,"p95_s_adjusted":3.6657787574997194,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":29,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.7604367004496922,"capability":43.05068960177584,"axes":{"calibration":77.50040036924202,"cost":78.3542548810428,"intelligence":8.600978834309654,"speed":72.8877498600171},"rank":201,"open_board_rank":158,"api_board_rank":null}},{"key":"h2o-lightning-31b","name":"H2O-Lightning-31B v1.0.7 (H2O.ai, Gemma-4-31B-it fine-tune, vLLM + h2oai overlay, one-engine FP8 profile, self-hosted)","board":"open","source_url":"https://huggingface.co/h2oai/h2o-lightning-31b","axes":{"intelligence":79.29803124474094,"calibration":89.37957627728649,"speed":89.5083860213906,"cost":56.02061699570665},"capability":{"score":84.33880376101371,"rank":11,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":7,"api_board_rank":null},"composite_score":75.72659612193334,"rank":22,"open_board_rank":14,"api_board_rank":null,"measured_in":"v1.6.39","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02924013904463419,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10147.2130346 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 727,319 input + 1,485 output tokens over 1,485 decisions = 490.777104377 tokens/decision. GPU-time cost USD 0.0292401390446342/1,000; old list-price cost USD 0.0444199393939394/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0292401390446342/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06450985649985341,"p50_s_adjusted":0.27901971299970685,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06450985649985341,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12567649844977494,"p95_s_adjusted":0.40135299689954984,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":73.43686411111796,"capability":85.16947336412102,"axes":{"calibration":90.18882631775024,"cost":50.57266115000792,"intelligence":80.15012041049178,"speed":89.8424830803211},"rank":12,"open_board_rank":3,"api_board_rank":null}},{"key":"h2o-lightning-31b-v1.3.4","name":"H2O-Lightning-31B v1.3.4 (@c35ed950)","board":"open","source_url":"https://huggingface.co/h2oai/h2o-lightning-31b","axes":{"intelligence":84.9859880104957,"calibration":90.72556650172525,"speed":90.34116786910276,"cost":60.193951589794054},"capability":{"score":87.85577725611049,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":79.2534131113009,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.13","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.13","measured_on":"2026-11-16","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.021226005886784718,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.0212260058867847/1,000; old list-price estimate USD 0.0321852518468771/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0212260058867847/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.052033289500059254,"p50_s_adjusted":0.25406657900011853,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.052033289500059254,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10693056890015652,"p95_s_adjusted":0.36386113780031304,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.13","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"h2o-lightning-4b","name":"H2O-Lightning-4B v1.1 (H2O.ai, Qwen3.5-4B fine-tune, stock vLLM + open shim)","board":"open","source_url":"https://huggingface.co/h2oai/h2o-lightning-4b","axes":{"intelligence":60.03839906196252,"calibration":89.95399367673842,"speed":92.37207479774074,"cost":78.58891530847164},"capability":{"score":74.99619636935047,"rank":66,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":49,"api_board_rank":null},"composite_score":77.93772981332452,"rank":6,"open_board_rank":3,"api_board_rank":null,"measured_in":"v1.6.1","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-04","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005172527675780014,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 854,867 input + 0 output tokens over 1,624 decisions = 526.395935961 tokens/decision. GPU-time cost USD 0.00517252767578001/1,000; old list-price cost USD 0.0157918780788177/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00517252767578001/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.031098264502361417,"p50_s_adjusted":0.21219652900472283,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.031098264502361417,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06146593304001726,"p95_s_adjusted":0.2729318660800345,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":73.82224524014528,"capability":74.99619636935047,"axes":{"calibration":89.95399367673842,"cost":64.04698652895101,"intelligence":60.03839906196252,"speed":92.61084474720658},"rank":10,"open_board_rank":2,"api_board_rank":null}},{"key":"h2o-lightning-4b-v1.4.1","name":"H2O-Lightning-4B v1.4.1","board":"open","source_url":"https://huggingface.co/h2oai/h2o-lightning-4b","axes":{"intelligence":61.544377442767995,"calibration":87.79935625372364,"speed":93.86173474461954,"cost":83.2204883200473},"capability":{"score":74.67186684824583,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":79.50976866203804,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.7.13","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.13","measured_on":"2026-10-27","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Measured on the 27 Oct 2026 draw and shown on the reference scale by anchor equating with four strong anchor systems (Intelligence -2.9, Calibration -1.9 points). Not ranked yet: its top-five places lie within the uncertainty of that shift (95% interval -5.9 to +1.1 Intelligence points). Ranking follows when more anchor runs narrow the interval.","price":{"kind":"estimate","usd_per_1000_decisions":0.003625075498997057,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 547,826 input + 1,489 output tokens over 1,489 decisions = 368.915379449 tokens/decision. GPU-time cost USD 0.00362507549899706/1,000; old list-price cost USD 0.0111874613834788/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00362507549899706/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.019854689999988295,"p50_s_adjusted":0.18970937999997659,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.019854689999988295,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03331976900000768,"p95_s_adjusted":0.21663953800001534,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.13","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"hopper","name":"Hopper","board":"open","source_url":"https://huggingface.co/HopitAI/hopper","axes":{"intelligence":33.59530941839744,"calibration":88.80153061371698,"speed":90.6039582415622,"cost":76.79820577848395},"capability":{"score":61.19842001605721,"rank":165,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":133,"api_board_rank":null},"composite_score":27.745611075479957,"rank":196,"open_board_rank":156,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005934623072864991,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 980,819 input + 0 output tokens over 1,624 decisions = 603.952586207 tokens/decision. GPU-time cost USD 0.00593462307286499/1,000; old list-price cost USD 0.0181185775862069/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00593462307286499/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.056871714536100626,"p50_s_adjusted":0.2637434290722013,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.056871714536100626,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08996527364943174,"p95_s_adjusted":0.3299305472988635,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":22,"n_over_limit_answered_ok":22,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":28.852652318563656,"capability":62.40936838189898,"axes":{"calibration":90.07902805935959,"cost":62.25627699896333,"intelligence":34.73970870443837,"speed":91.11434937226889},"rank":104,"open_board_rank":72,"api_board_rank":null}},{"key":"hopper-12b-trained","name":"Hopper 12B trained (gemma-4-12B-it frozen + LoRA r32 unmerged, one forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/HopitAI/hopper","axes":{"intelligence":51.6576814732076,"calibration":84.63452107883388,"speed":83.63977704006717,"cost":66.55744455639032},"capability":{"score":68.14610127602074,"rank":119,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":92,"api_board_rank":null},"composite_score":68.78245633029094,"rank":107,"open_board_rank":83,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01302422373612839,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 854,727 input + 0 output tokens over 1,477 decisions = 578.69126608 tokens/decision. GPU-time cost USD 0.0130242237361284/1,000; old list-price cost USD 0.0289345633039946/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0130242237361284/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.18336594451102428,"p50_s_adjusted":0.5167318890220486,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.18336594451102428,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.34353042687289415,"p95_s_adjusted":0.8370608537457883,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":65.78419365065008,"capability":68.14610127602074,"axes":{"calibration":84.63452107883388,"cost":56.15749203571245,"intelligence":51.6576814732076,"speed":84.57301814857726},"rank":47,"open_board_rank":26,"api_board_rank":null}},{"key":"hyland-hs1","name":"HS1 0.1.2 (Hyland)","board":"api","source_url":"https://www.hyland.com","axes":{"intelligence":52.865450899914414,"calibration":75.8316875668352,"speed":93.51379976299526,"cost":49.09977911142268},"capability":{"score":64.3485692333748,"rank":135,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":28},"composite_score":61.068034493527705,"rank":128,"open_board_rank":null,"api_board_rank":30,"measured_in":"v1.6.36","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.04973636363636364,"basis":"ESTIMATE, v1.5-M2 base-model reference rule. The customer states a benchmark tariff of USD 0.056 per million input tokens, output free (one NVIDIA L40S, AWS on-demand); there is no public price page for it, so it is not a published list price. Under the method, the base-model reference price applies as an ESTIMATE for this vendor-hosted row, as for attest-31b-v1.0: base model google/gemma-4-26b-a4b-it through the frozen pricing_v15 module (sha256 1f4d1f7b1ce97ab27b938043444b9c2757e4a9dbecc799a803122588401f5bf1) reference() -> google/gemma-4-26b-a4b-it on OpenRouter, USD 0.09 per million input tokens, USD 0.30 per million output tokens. The customer-stated tariff is shown for information and is not used for the Cost axis. Applied to this row's own provider-reported tokens, first answer per item (2,012,340 input and 0 output tokens over the 1,500 items; 820,650 input and 0 output tokens over the 1,485 items of the common cost basis; the 15 very long items count for no row's cost, whether the row refused or answered them). Estimated, not a vendor bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them in full)."},"latency":{"p50_s_raw":0.19862511445535347,"p50_s_adjusted":0.19862511445535347,"adjustment":"none (API)","p50_s_measured":0.19862511445535347,"latency_basis":"api-end-to-end","p95_s_raw":0.2241743211168796,"p95_s_adjusted":0.2241743211168796,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":61.450719939462154,"capability":65.03939854649335,"axes":{"calibration":76.49930909839343,"cost":49.09977911142268,"intelligence":53.57948799459325,"speed":93.74328417303443},"rank":56,"open_board_rank":null,"api_board_rank":24}},{"key":"imajev-2b","name":"Imajev-2B (mohit67890, Qwen3.5-2B LoRA)","board":"open","source_url":"https://huggingface.co/mohit67890/imajev-2b","axes":{"intelligence":21.160522003324633,"calibration":86.77653179414553,"speed":81.2429913982843,"cost":67.34930355763724},"capability":{"score":53.96852689873508,"rank":207,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":173,"api_board_rank":null},"composite_score":8.336534540273085,"rank":252,"open_board_rank":204,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.012256220002939449,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 3,010,517 input + 0 output tokens over 1,485 decisions = 2027.28417508 tokens/decision. GPU-time cost USD 0.0122562200029394/1,000; old list-price cost USD 0.0405456835016835/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0122562200029394/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.2978108120005345,"p50_s_adjusted":0.745621624001069,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2978108120005345,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.42867729024728757,"p95_s_adjusted":1.0073545804945752,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":7.92834166805135,"capability":53.96852689873508,"axes":{"calibration":86.77653179414553,"cost":51.76166122103225,"intelligence":21.160522003324633,"speed":81.55302418750827},"rank":177,"open_board_rank":138,"api_board_rank":null}},{"key":"imajev-4b-rtx5090-a2","name":"Imajev-4B (RTX 5090)","board":"open","source_url":"https://github.com/fstandhartinger/jevbench/issues/110","axes":{"intelligence":34.62714018309005,"calibration":89.1627567462182,"speed":90.38312320077796,"cost":77.80651098778299},"capability":{"score":61.89494846465412,"rank":155,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":123,"api_board_rank":null},"composite_score":29.970850718248183,"rank":193,"open_board_rank":153,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005492663166795393,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 907,776 input + 0 output tokens over 1,624 decisions = 558.975369458 tokens/decision. GPU-time cost USD 0.00549266316679539/1,000; old list-price cost USD 0.0167692610837438/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00549266316679539/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07211481148260646,"p50_s_adjusted":0.29422962296521293,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07211481148260646,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08058627708815037,"p95_s_adjusted":0.31117255417630074,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":28.655751132526984,"capability":61.89494846465412,"axes":{"calibration":89.1627567462182,"cost":63.26458220826236,"intelligence":34.62714018309005,"speed":90.51878000442085},"rank":105,"open_board_rank":73,"api_board_rank":null}},{"key":"imajev-9b","name":"Imajev-9B (mohit67890, Qwen3.5-9B LoRA)","board":"open","source_url":"https://huggingface.co/mohit67890/imajev-9b","axes":{"intelligence":36.7327337780622,"calibration":88.10935553643324,"speed":79.23152050102456,"cost":50.455376685541566},"capability":{"score":62.42104465724772,"rank":153,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":121,"api_board_rank":null},"composite_score":30.400650379076136,"rank":191,"open_board_rank":151,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.044821609112740766,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 3,010,517 input + 0 output tokens over 1,485 decisions = 2027.28417508 tokens/decision. GPU-time cost USD 0.0448216091127408/1,000; old list-price cost USD 0.202728417508417/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0448216091127408/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.39089938551478554,"p50_s_adjusted":0.9317987710295711,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.39089938551478554,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.5654656185419299,"p95_s_adjusted":1.2809312370838597,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":9.790864242743115,"capability":62.42104465724772,"axes":{"calibration":88.10935553643324,"cost":30.792561090951693,"intelligence":36.7327337780622,"speed":79.48842001125843},"rank":172,"open_board_rank":133,"api_board_rank":null}},{"key":"instinct","name":"Instinct (ZooWork, Qwen3.8-27B)","board":"api","source_url":"https://github.com/fstandhartinger/jevbench/issues/69","axes":{"intelligence":47.77,"calibration":93.17,"speed":87.14,"cost":64.41},"capability":{"score":70.47,"rank":100,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":25},"composite_score":62.23,"rank":127,"open_board_rank":null,"api_board_rank":29,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.015355329949238578,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.03/M input, 0/M output (spec instinct.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.296,"p50_s_adjusted":0.296,"adjustment":"none (API)","p50_s_measured":0.296,"latency_basis":"api-end-to-end","p95_s_raw":0.653,"p95_s_adjusted":0.653,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":62.24,"capability":70.47,"axes":{"calibration":93.17,"cost":64.41,"intelligence":47.77,"speed":87.2},"rank":55,"open_board_rank":null,"api_board_rank":23}},{"key":"instinct-dual-4b","name":"Instinct Dual 4B","board":"api","source_url":"https://github.com/fstandhartinger/jevbench/issues/123","axes":{"intelligence":28.34,"calibration":86.08,"speed":88.56,"cost":75.03},"capability":{"score":57.21,"rank":180,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":32},"composite_score":17.97,"rank":222,"open_board_rank":null,"api_board_rank":44,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.006799306260575296,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.01/M input, 0/M output (spec instinct-dual-4b.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.232,"p50_s_adjusted":0.232,"adjustment":"none (API)","p50_s_measured":0.232,"latency_basis":"api-end-to-end","p95_s_raw":0.601,"p95_s_adjusted":0.601,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.97,"capability":57.21,"axes":{"calibration":86.08,"cost":75.03,"intelligence":28.34,"speed":88.6},"rank":142,"open_board_rank":null,"api_board_rank":36}},{"key":"iota-decide-4b","name":"iota-decide-4b (Macrocosmos, Clef joint-head 4B, self-hosted)","board":"open","source_url":"https://huggingface.co/macrocosm-os/iota-decide-4b","axes":{"intelligence":29.414053145614563,"calibration":84.8346261602328,"speed":87.14949384308518,"cost":77.51227090165808},"capability":{"score":57.12433965292368,"rank":184,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":150,"api_board_rank":null},"composite_score":19.7303928088595,"rank":215,"open_board_rank":172,"api_board_rank":null,"measured_in":"v1.6.38","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005618119465543728,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 849,038 input + 0 output tokens over 1,485 decisions = 571.742760943 tokens/decision. GPU-time cost USD 0.00561811946554373/1,000; old list-price cost USD 0.0171522828282828/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00561811946554373/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12183865150007023,"p50_s_adjusted":0.3936773030001405,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12183865150007023,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16983880055001918,"p95_s_adjusted":0.48967760110003833,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.937474020302744,"capability":57.12433965292368,"axes":{"calibration":84.8346261602328,"cost":62.97034212213745,"intelligence":29.414053145614563,"speed":87.4670144658746},"rank":135,"open_board_rank":100,"api_board_rank":null}},{"key":"iota-decide-4b-v0.3","name":"iota-decide-4b-v0.3 (Macrocosmos, letter-readout 4B, self-hosted)","board":"open","source_url":"https://huggingface.co/macrocosm-os/iota-decide-4b-v0.3","axes":{"intelligence":52.532639982463,"calibration":84.75645357480421,"speed":87.46776808261981,"cost":78.86867492076232},"capability":{"score":68.64454677863361,"rank":116,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":89,"api_board_rank":null},"composite_score":72.7982392443752,"rank":66,"open_board_rank":50,"api_board_rank":null,"measured_in":"v1.7.10","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.10","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005062645321094913,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 765,092 input + 0 output tokens over 1,485 decisions = 515.213468013 tokens/decision. GPU-time cost USD 0.00506264532109491/1,000; old list-price cost USD 0.015456404040404/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00506264532109491/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11314048350004668,"p50_s_adjusted":0.3762809670000934,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11314048350004668,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16305699360005405,"p95_s_adjusted":0.4761139872001081,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.10","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"iota-decide-4b-v0.4","name":"iota-decide-4b-v0.4 (Macrocosmos, letter-readout 4B, self-hosted)","board":"open","source_url":"https://huggingface.co/macrocosm-os/iota-decide-4b-v0.4","axes":{"intelligence":57.322898895925036,"calibration":85.90496615662252,"speed":92.23793413696737,"cost":83.01507877958139},"capability":{"score":71.61393252627377,"rank":87,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":64,"api_board_rank":null},"composite_score":76.96253989391107,"rank":16,"open_board_rank":11,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0036826804870835747,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 558,044 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00368268048708357/1,000; old list-price estimate USD 0.0112433310946944/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00368268048708357/1,000. Estimated, not a bill. USD 0.00368268049 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00420412617 over all 1,500 items)."},"latency":{"p50_s_raw":0.03286642200001211,"p50_s_adjusted":0.2157328440000242,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.03286642200001211,"latency_basis":"api-end-to-end","p95_s_raw":0.06343959554999401,"p95_s_adjusted":0.27687919109998804,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jade","name":"JADE (Qwen3.8-27B LoRA)","board":"open","source_url":"https://huggingface.co/theunnecessarythings/JADE","axes":{"intelligence":47.93439876939806,"calibration":50.5645403785324,"speed":85.09218748397188,"cost":61.10443159149903},"capability":{"score":49.24946957396523,"rank":221,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":185,"api_board_rank":null},"composite_score":53.46928254382284,"rank":138,"open_board_rank":108,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.019793333716611017,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 567,713 input + 1,108 output tokens over 1,477 decisions = 385.11916046 tokens/decision. GPU-time cost USD 0.019793333716611/1,000; old list-price cost USD 0.16368548408937/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.019793333716611/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13592451099248137,"p50_s_adjusted":0.42184902198496277,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13592451099248137,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2919392841402441,"p95_s_adjusted":0.7338785682804883,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.194459825016597,"capability":49.24946957396523,"axes":{"calibration":50.5645403785324,"cost":33.57969498589053,"intelligence":47.93439876939806,"speed":85.53222225487117},"rank":131,"open_board_rank":96,"api_board_rank":null}},{"key":"janus-4b","name":"janus 4B (Icarus AI / cmxu, Qwen3.5-4B + LoRA r64 and pointer decision head)","board":"open","source_url":"https://huggingface.co/cmxu/janus-4b","axes":{"intelligence":37.57770128683752,"calibration":88.12743013642095,"speed":92.71898033409317,"cost":78.95653992094968},"capability":{"score":62.85256571162924,"rank":148,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":117,"api_board_rank":null},"composite_score":36.791451915443076,"rank":182,"open_board_rank":144,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005028618274386871,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 831,083 input + 0 output tokens over 1,624 decisions = 511.750615764 tokens/decision. GPU-time cost USD 0.00502861827438687/1,000; old list-price cost USD 0.0153525184729064/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00502861827438687/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03142766794189811,"p50_s_adjusted":0.2128553358837962,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03142766794189811,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05059936083853242,"p95_s_adjusted":0.2511987216770648,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":35.18351981268845,"capability":62.85256571162924,"axes":{"calibration":88.12743013642095,"cost":64.41461114142905,"intelligence":37.57770128683752,"speed":93.17471817821269},"rank":93,"open_board_rank":63,"api_board_rank":null}},{"key":"jeb-35b-a3b","name":"JEB-35B-A3B (szybkie.ai, Qwen3.6-35B-A3B LoRA)","board":"open","source_url":"https://huggingface.co/szybkie-ai/jeb-35b-a3b","axes":{"intelligence":33.22856158102424,"calibration":89.98387597336676,"speed":92.86051128678476,"cost":72.29693310055532},"capability":{"score":61.6062187771955,"rank":162,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":130,"api_board_rank":null},"composite_score":26.844904879645952,"rank":200,"open_board_rank":159,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008383696796870014,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 781,043 input + 1,485 output tokens over 1,485 decisions = 526.954882155 tokens/decision. GPU-time cost USD 0.00838369679687001/1,000; old list-price cost USD 0.0798932323232323/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00838369679687001/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03574157250113785,"p50_s_adjusted":0.2214831450022757,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03574157250113785,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.041836407268419855,"p95_s_adjusted":0.2336728145368397,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.300802214285724,"capability":61.6062187771955,"axes":{"calibration":89.98387597336676,"cost":42.92470023353102,"intelligence":33.22856158102424,"speed":92.99325391268877},"rank":145,"open_board_rank":109,"api_board_rank":null}},{"key":"jebadiah-27b","name":"Jebadiah 27B (Frontier Infra, Qwen3.8-27B LoRA)","board":"open","source_url":"https://huggingface.co/frontier-infra/jebadiah-27b","axes":{"intelligence":57.4596907179922,"calibration":83.00593897276424,"speed":85.07834149803253,"cost":57.46143151439629},"capability":{"score":70.23281484537821,"rank":103,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":77,"api_board_rank":null},"composite_score":68.25043421482714,"rank":113,"open_board_rank":89,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.026178953569491424,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 750,854 input + 1,477 output tokens over 1,477 decisions = 509.364251862 tokens/decision. GPU-time cost USD 0.0261789535694914/1,000; old list-price cost USD 0.21651298578199/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0261789535694914/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.156828826511628,"p50_s_adjusted":0.463657653023256,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.156828826511628,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.25991794824367387,"p95_s_adjusted":0.6698358964873478,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":19.255307672764186,"capability":70.3119788331729,"axes":{"calibration":83.03844121699153,"cost":29.935481526523958,"intelligence":57.58551644935426,"speed":85.83700487110255},"rank":134,"open_board_rank":99,"api_board_rank":null}},{"key":"jebadiah-9b-v2","name":"Jebadiah 9B v2 (Frontier Infra, Qwen3.5-9B LoRA)","board":"open","source_url":"https://huggingface.co/frontier-infra/jebadiah-9b-v2","axes":{"intelligence":30.5871791271758,"calibration":83.31146464049283,"speed":91.458678539542,"cost":68.53174752675467},"capability":{"score":56.949321883834315,"rank":186,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":152,"api_board_rank":null},"composite_score":21.316898270250412,"rank":211,"open_board_rank":169,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011192877459801475,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 750,303 input + 1,485 output tokens over 1,485 decisions = 506.254545455 tokens/decision. GPU-time cost USD 0.0111928774598015/1,000; old list-price cost USD 0.0506754545454545/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0111928774598015/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05506055150181055,"p50_s_adjusted":0.2601211030036211,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05506055150181055,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06238096526358276,"p95_s_adjusted":0.2747619305271655,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.78430461690739,"capability":56.949321883834315,"axes":{"calibration":83.31146464049283,"cost":48.85607041331758,"intelligence":30.5871791271758,"speed":91.57148361254983},"rank":137,"open_board_rank":102,"api_board_rank":null}},{"key":"jebadiah-9b-v2-1","name":"Jebadiah 9B v2.1 (Frontier Infra, Qwen3.5-9B LoRA)","board":"open","source_url":"https://huggingface.co/frontier-infra/jebadiah-9b-v2-1","axes":{"intelligence":37.15024789587638,"calibration":89.67673072100749,"speed":90.97576909567059,"cost":68.53174752675467},"capability":{"score":63.41348930844194,"rank":145,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":114,"api_board_rank":null},"composite_score":34.69183418142899,"rank":183,"open_board_rank":145,"api_board_rank":null,"measured_in":"v1.6.25","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011192877459801475,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 750,303 input + 1,485 output tokens over 1,485 decisions = 506.254545455 tokens/decision. GPU-time cost USD 0.0111928774598015/1,000; old list-price cost USD 0.0506754545454545/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0111928774598015/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0503088840014243,"p50_s_adjusted":0.25061776800284863,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0503088840014243,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08436070300049323,"p95_s_adjusted":0.3187214060009864,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":30.34536701293529,"capability":63.41348930844194,"axes":{"calibration":89.67673072100749,"cost":48.85607041331758,"intelligence":37.15024789587638,"speed":91.39972504887733},"rank":100,"open_board_rank":68,"api_board_rank":null}},{"key":"jeetu-decider-2b","name":"Jeetu Decider 2B","board":"open","source_url":"https://huggingface.co/Jeetu456/decider-2b","axes":{"intelligence":10.191329118370533,"calibration":73.13553830274921,"speed":90.89924129769977,"cost":85.8779611353149},"capability":{"score":41.66343371055987,"rank":248,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":211,"api_board_rank":null},"composite_score":1.23608490981135,"rank":270,"open_board_rank":219,"api_board_rank":null,"measured_in":"v1.7.2","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.2","measured_on":"2026-10-25","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0029562055555555564,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 726,138 input + 0 output tokens over 1,485 decisions = 488.981818182 tokens/decision. GPU-time cost USD 0.00295620555555556/1,000; old list-price cost USD 0.00977963636363636/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00295620555555556/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.060729330999947706,"p50_s_adjusted":0.27145866199989543,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.060729330999947706,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07474149694996866,"p95_s_adjusted":0.2994829938999373,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.2","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jeff_1_0_large","name":"Jeff-1.0-Large","board":"open","source_url":"https://huggingface.co/jgeuter/Jeff-1.0-Large","axes":{"intelligence":71.89886397161271,"calibration":91.10757097649757,"speed":88.47496650745126,"cost":54.523640029233874},"capability":{"score":81.50321747405513,"rank":27,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":20,"api_board_rank":null},"composite_score":73.35712724451902,"rank":54,"open_board_rank":39,"api_board_rank":null,"measured_in":"v1.6.2","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-09","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03280036418109335,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 810,973 input + 0 output tokens over 1,479 decisions = 548.325219743 tokens/decision. GPU-time cost USD 0.0328003641810933/1,000; old list-price cost USD 0.0493492697768763/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0328003641810933/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.08351965248584747,"p50_s_adjusted":0.31703930497169497,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08351965248584747,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.14905789233744146,"p95_s_adjusted":0.4481157846748829,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.59329511867139,"capability":81.50321747405513,"axes":{"calibration":91.10757097649757,"cost":49.20157807658991,"intelligence":71.89886397161271,"speed":88.82411852274376},"rank":32,"open_board_rank":15,"api_board_rank":null}},{"key":"jeff_1_1_large","name":"Jeff-1.1-Large (jgeuter)","board":"open","source_url":"https://huggingface.co/jgeuter/Jeff-1.1-Large","axes":{"intelligence":81.60494681787081,"calibration":90.10053914936476,"speed":72.19753970140297,"cost":58.304016843105764},"capability":{"score":85.85274298361779,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":73.58996625434015,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.7.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.14","measured_on":"2026-10-27","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Measured on the 27 Oct 2026 draw and shown on the reference scale by anchor equating with four strong anchor systems (Intelligence -2.9, Calibration -1.9 points). Not ranked yet: its top-five places lie within the uncertainty of that shift (95% interval -5.9 to +1.1 Intelligence points). Ranking follows when more anchor runs narrow the interval.","price":{"kind":"estimate","usd_per_1000_decisions":0.02453952235507212,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 589,108 input + 21,722 output tokens over 1,489 decisions = 410.228341169 tokens/decision. GPU-time cost USD 0.0245395223550721/1,000; old list-price cost USD 0.0405676292813969/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0245395223550721/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1056369350001205,"p50_s_adjusted":0.361273870000241,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1056369350001205,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":8.269099666149986,"p95_s_adjusted":16.68819933229997,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.14","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jeff_1_1_medium","name":"Jeff-1.1-Medium (jgeuter)","board":"open","source_url":"https://huggingface.co/jgeuter/Jeff-1.1-Medium","axes":{"intelligence":74.49393760932514,"calibration":89.2826256096074,"speed":73.9814920153467,"cost":66.15113749144632},"capability":{"score":81.88828160946628,"rank":21,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":15,"api_board_rank":null},"composite_score":75.10595924834655,"rank":28,"open_board_rank":18,"api_board_rank":null,"measured_in":"v1.7.8","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.8","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.013436786510251734,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 814,647 input + 71,931 output tokens over 1,485 decisions = 597.022222222 tokens/decision. GPU-time cost USD 0.0134367865102517/1,000; old list-price cost USD 0.0346949494949495/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0134367865102517/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06713257949991203,"p50_s_adjusted":0.2842651589998241,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06713257949991203,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":6.957296431550037,"p95_s_adjusted":14.064592863100074,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.8","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jeff_1_1_qwen_4b","name":"Jeff-1.1-Qwen-4B (jgeuter)","board":"open","source_url":"https://huggingface.co/jgeuter/Jeff-1.1-Qwen-4B","axes":{"intelligence":65.284729247328,"calibration":89.24636404689907,"speed":78.37104008757272,"cost":80.68251138779327},"capability":{"score":77.26554664711354,"rank":54,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":40,"api_board_rank":null},"composite_score":77.4045764244785,"rank":11,"open_board_rank":6,"api_board_rank":null,"measured_in":"v1.7.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.14","measured_on":"2026-10-27","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004404699520370065,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 598,499 input + 68,954 output tokens over 1,489 decisions = 448.255876427 tokens/decision. GPU-time cost USD 0.00440469952037007/1,000; old list-price cost USD 0.0190047481531229/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00440469952037007/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.039202700000032564,"p50_s_adjusted":0.22840540000006512,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.039202700000032564,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":3.110368112699962,"p95_s_adjusted":6.3707362253999245,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.14","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jeffan-31b-1710","name":"Jeffan-31b-1710","board":"open","source_url":"https://huggingface.co/tattrongvu/Jeffan-31b-1710","axes":{"intelligence":71.63230144667592,"calibration":91.73503889029561,"speed":84.65261351124272,"cost":51.40971770392302},"capability":{"score":81.68367016848578,"rank":24,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":18,"api_board_rank":null},"composite_score":71.26859587879677,"rank":84,"open_board_rank":66,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0416558572638797,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 13044.9643397 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 1,334,761 input + 0 output tokens over 1,485 decisions = 898.828956229 tokens/decision. GPU-time cost USD 0.0416558572638797/1,000; old list-price cost USD 0.0808946060606061/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0416558572638797/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12250980800013167,"p50_s_adjusted":0.39501961600026336,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12250980800013167,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3586007160500799,"p95_s_adjusted":0.8672014321001599,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jeffan-4b-2010","name":"Jeffan-4B-2010","board":"open","source_url":"https://huggingface.co/tattrongvu/Jeffan-4B-2010","axes":{"intelligence":40.09995458474205,"calibration":87.9034951152282,"speed":93.04878911983016,"cost":78.67753572713382},"capability":{"score":64.00172484998512,"rank":139,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":109,"api_board_rank":null},"composite_score":43.04448867574887,"rank":164,"open_board_rank":130,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005137464206465065,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 776,399 input + 0 output tokens over 1,485 decisions = 522.827609428 tokens/decision. GPU-time cost USD 0.00513746420646507/1,000; old list-price cost USD 0.0156848282828283/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00513746420646507/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02829001850000168,"p50_s_adjusted":0.20658003700000335,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02829001850000168,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04495068749999973,"p95_s_adjusted":0.23990137499999947,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jev-1.13.0","name":"Jev 1.13.0 (TypeSafe AI)","board":"reference","source_url":"https://docs.typesafe.ai","axes":{"intelligence":63.62075784024779,"calibration":90.63220203152369,"speed":91.21186387733223,"cost":54.733842199536014},"capability":{"score":77.12647993588574,"rank":56,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":71.44711653093671,"rank":83,"open_board_rank":null,"api_board_rank":18,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.03227542180094787,"basis":"TypeSafe list tariff USD 0.042 per 1M input tokens, output free; measured on the v1.6.1 common cost basis (all items except those outside the Jev reference input range) from this run's token usage"},"latency":{"p50_s_raw":0.24289856851100922,"p50_s_adjusted":0.24289856851100922,"adjustment":"none (API)","p50_s_measured":0.24289856851100922,"latency_basis":"api-end-to-end","p95_s_raw":0.31145023070275785,"p95_s_adjusted":0.31145023070275785,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":71.49118639284514,"capability":77.12647993588574,"axes":{"calibration":90.63220203152369,"cost":54.733842199536014,"intelligence":63.62075784024779,"speed":91.49989299046987},"rank":18,"open_board_rank":null,"api_board_rank":12}},{"key":"jev-1.14.0","name":"Jev 1.14.0 (TypeSafe AI)","board":"api","source_url":"https://docs.typesafe.ai/models","axes":{"intelligence":59.942981848430136,"calibration":90.08195706553009,"speed":90.98627348971067,"cost":54.815314343766296},"capability":{"score":75.01246945698011,"rank":65,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":16},"composite_score":70.1541854578216,"rank":95,"open_board_rank":null,"api_board_rank":22,"measured_in":"v1.6.27","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.032074226262626265,"basis":"TypeSafe AI public list price, USD 0.042/M input, output free: the \"Current models\" table of TypeSafe's models page (docs.typesafe.ai/models, fetched 16 Oct 2026 22:37:44Z, saved copy sha256 936da179...) lists jev-1.14.0 at \"$42 / $0.042\" per Btok / per Mtok, the same price as jev-1.13.0, and says \"You pay per input token. Output tokens are free.\"; Jev 1.14 is \"available from October 14, 2026\". Counted as the manufacturer's standard launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap, d1-ultra, Sage 1.4, Nimble, kettei and Celeris-1 Decision), through a scoped re-freeze of the frozen pricing module that adds this one key and nothing else, with a zero-difference proof for every other registry entry; jev-1.13.0 is priced through the same launch-list set at the same tariff. On the same page the alias jev-latest (\"the most recent stable, official release\") still points to jev-1.13.0, and jev-preview (\"the most recent release, official or not\") points to jev-1.14.0; the versioned id is listed with its price, accepted by the API and was the served model on every row. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (1,134,053 input tokens over the 1,485 answered items; the 71,566 reported output tokens are free under the tariff), divided by the 1,485 items of the common cost basis the scorer uses; the 15 requests refused with HTTP 400 report no usage and count as zero tokens and as wrong answers (TypeSafe's docs do not say whether refused requests are billed). USD 0.032074 per 1,000 decisions on the common basis (USD 0.031753 over all 1,500 items), 0.497x the official cost cap. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.032074 per 1,000 decisions (0.497x the official cost cap USD 0.064595), against USD 0.031753 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.22666933800792322,"p50_s_adjusted":0.22666933800792322,"adjustment":"none (API)","p50_s_measured":0.22666933800792322,"latency_basis":"api-end-to-end","p95_s_raw":0.35154414957505653,"p95_s_adjusted":0.35154414957505653,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.1509594948736,"capability":75.01246945698011,"axes":{"calibration":90.08195706553009,"cost":54.815314343766296,"intelligence":59.942981848430136,"speed":90.96457246082724},"rank":26,"open_board_rank":null,"api_board_rank":16}},{"key":"jev-27b-autotrust","name":"AutoTrust-Decider-Pro-27B (AutoTrust, Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/autotrust/JEV-27B","axes":{"intelligence":56.828664559360675,"calibration":90.89967561034447,"speed":86.4682823414372,"cost":58.5653810087091},"capability":{"score":73.86417008485257,"rank":71,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":52,"api_board_rank":null},"composite_score":69.88393575514445,"rank":99,"open_board_rank":77,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.024052153474629835,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 689,734 input + 1,477 output tokens over 1,477 decisions = 467.983073798 tokens/decision. GPU-time cost USD 0.0240521534746298/1,000; old list-price cost USD 0.199132890995261/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0240521534746298/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11307674099225551,"p50_s_adjusted":0.37615348198451104,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11307674099225551,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.22476207326253642,"p95_s_adjusted":0.5995241465250728,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.291230595611577,"capability":73.86417008485257,"axes":{"calibration":90.89967561034447,"cost":31.02571003473804,"intelligence":56.828664559360675,"speed":86.89503266505372},"rank":122,"open_board_rank":88,"api_board_rank":null}},{"key":"jev-local","name":"jev-local (Qwen3.5-9B)","board":"open","source_url":"https://github.com/us/jev-local","axes":{"intelligence":41.85041350559534,"calibration":73.06634984647151,"speed":72.89461160481758,"cost":78.54608301342061},"capability":{"score":57.458381676033426,"rank":179,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":147,"api_board_rank":null},"composite_score":43.76548052831858,"rank":159,"open_board_rank":126,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0051895603433263346,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 379,568 input + 1,624 output tokens over 1,624 decisions = 234.724137931 tokens/decision. GPU-time cost USD 0.00518956034332633/1,000; old list-price cost USD 0.0235224137931034/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00518956034332633/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.47121412493288517,"p50_s_adjusted":1.0924282498657703,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.47121412493288517,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.2752600665669886,"p95_s_adjusted":4.700520133133978,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":41.328564334669615,"capability":57.458381676033426,"axes":{"calibration":73.06634984647151,"cost":58.85554343293177,"intelligence":41.85041350559534,"speed":75.55955847735143},"rank":79,"open_board_rank":52,"api_board_rank":null}},{"key":"jev-omni","name":"Jev-Omni (akhilaaa3, Gemma-4-12B merged)","board":"open","source_url":"https://huggingface.co/akhilaaa3/Jev-Omni","axes":{"intelligence":55.530225828826175,"calibration":87.04450672835743,"speed":84.73829853475769,"cost":66.45108408007525},"capability":{"score":71.2873662785918,"rank":90,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":67,"api_board_rank":null},"composite_score":70.98959059698258,"rank":87,"open_board_rank":68,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.013130981737865803,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 947,498 input + 0 output tokens over 1,624 decisions = 583.434729064 tokens/decision. GPU-time cost USD 0.0131309817378658/1,000; old list-price cost USD 0.029171736453202/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0131309817378658/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1615908145904541,"p50_s_adjusted":0.4731816291809082,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1615908145904541,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2799051289446652,"p95_s_adjusted":0.7098102578893304,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.74758858028281,"capability":71.2873662785918,"axes":{"calibration":87.04450672835743,"cost":56.05113155939739,"intelligence":55.530225828826175,"speed":85.43173201875288},"rank":40,"open_board_rank":20,"api_board_rank":null}},{"key":"jev-qwen3.5-9b-base-nvfp4","name":"JEV Qwen3.5-9B Base NVFP4","board":"open","source_url":"https://huggingface.co/WIlfLin/JEV-Qwen3.5-9B-Base-NVFP4","axes":{"intelligence":29.251131763478327,"calibration":85.03288206087805,"speed":93.41131234211913,"cost":47.592235403520306},"capability":{"score":57.14200691217819,"rank":183,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":149,"api_board_rank":null},"composite_score":15.970464603997094,"rank":232,"open_board_rank":186,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.05583743842364531,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (proxy tokens, I-2): exact base-model market reference"},"latency":{"p50_s_raw":0.02305087901186198,"p50_s_adjusted":0.19610175802372395,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02305087901186198,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04124045012868009,"p95_s_adjusted":0.23248090025736018,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.97555620862577,"capability":57.14200691217819,"axes":{"calibration":85.03288206087805,"cost":47.592235403520306,"intelligence":29.251131763478327,"speed":93.62779491136789},"rank":154,"open_board_rank":116,"api_board_rank":null}},{"key":"jevact","name":"JevAct (einptein, jev1-2b-v2)","board":"api","source_url":"https://github.com/fstandhartinger/jevbench/issues/66","axes":{"intelligence":10.1,"calibration":47.44,"speed":76.36,"cost":68.56},"capability":{"score":28.77,"rank":272,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":38},"composite_score":1.11,"rank":271,"open_board_rank":null,"api_board_rank":52,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011167487684729064,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: v1.5 measured-input proxy; no exact base-model floor applies"},"latency":{"p50_s_raw":0.404,"p50_s_adjusted":0.808,"adjustment":"x2 (demo)","p50_s_measured":0.404,"latency_basis":"demo-endpoint-x2","p95_s_raw":1.431,"p95_s_adjusted":2.861,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":1.11,"capability":28.77,"axes":{"calibration":47.44,"cost":68.56,"intelligence":10.1,"speed":76.45},"rank":198,"open_board_rank":null,"api_board_rank":43}},{"key":"jevk5-v0.3-4b","name":"JevK5 v0.3 (4B)","board":"open","source_url":"https://huggingface.co/alibiserikbay/JevK5","axes":{"intelligence":38.53201254374527,"calibration":89.80312361577623,"speed":93.8287164149578,"cost":77.61552273233744},"capability":{"score":64.16756807976074,"rank":136,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":107,"api_board_rank":null},"composite_score":39.181248303247145,"rank":176,"open_board_rank":139,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.019413025584071875,"p50_s_adjusted":0.18882605116814374,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.019413025584071875,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03465702311368657,"p95_s_adjusted":0.21931404622737313,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":37.36493548751909,"capability":64.20716028123837,"axes":{"calibration":89.88230801873149,"cost":63.07359395281681,"intelligence":38.53201254374527,"speed":93.9472917853559},"rank":86,"open_board_rank":56,"api_board_rank":null}},{"key":"jevk5-v02","name":"JevK5 v0.2.0","board":"open","source_url":"https://github.com/allebee/jevk5","axes":{"intelligence":36.261123811123895,"calibration":89.38710082819252,"speed":91.22516918010956,"cost":77.61552273233744},"capability":{"score":62.824112319658205,"rank":150,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":118,"api_board_rank":null},"composite_score":33.60100126653793,"rank":184,"open_board_rank":146,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04432265050127171,"p50_s_adjusted":0.23864530100254341,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04432265050127171,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08301568527822382,"p95_s_adjusted":0.31603137055644764,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":32.09901620353128,"capability":62.824112319658205,"axes":{"calibration":89.38710082819252,"cost":63.07359395281681,"intelligence":36.261123811123895,"speed":91.56570599843724},"rank":97,"open_board_rank":66,"api_board_rank":null}},{"key":"jevone","name":"JevOne","board":"open","source_url":"https://huggingface.co/juspay/jev-one","axes":{"intelligence":45.955077428728934,"calibration":91.76581828544445,"speed":89.58480866139566,"cost":72.79744483400026},"capability":{"score":68.8604478570867,"rank":113,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":86,"api_board_rank":null},"composite_score":58.70681958482434,"rank":133,"open_board_rank":103,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008067738981245761,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 821,899 input + 1,624 output tokens over 1,624 decisions = 507.09544335 tokens/decision. GPU-time cost USD 0.00806773898124576/1,000; old list-price cost USD 0.101219088669951/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00806773898124576/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0652019064873457,"p50_s_adjusted":0.2804038129746914,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0652019064873457,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12120281010866166,"p95_s_adjusted":0.3924056202173233,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":31.14510331640707,"capability":68.8604478570867,"axes":{"calibration":91.76581828544445,"cost":39.84212731591276,"intelligence":45.955077428728934,"speed":89.83885080776159},"rank":98,"open_board_rank":67,"api_board_rank":null}},{"key":"jevora-9b-v1","name":"jevora-9b-v1 (kadirbekovvv / doxanocap, Qwen3.5-9B-Base derivative)","board":"open","source_url":"https://huggingface.co/kadirbekovvv/jevora-9b-v1","axes":{"intelligence":18.416504152111088,"calibration":55.43211615661533,"speed":90.40620976128842,"cost":null},"capability":{"score":36.92431015436321,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base and raises rather than estimating; one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 3948e3f1..., and the later re-freezes, which only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base (the module RAISES 'base model lacks an explicit market-listing decision'; the artifact's own metadata names this base); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens as reported by the author's in-process adapter (input 681,813, output_tokens 0)."},"latency":{"p50_s_raw":0.06486371951177716,"p50_s_adjusted":0.27972743902355435,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06486371951177716,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0877848407253623,"p95_s_adjusted":0.3255696814507246,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":36.99774224668246,"axes":{"calibration":55.45315460989178,"cost":null,"intelligence":18.54232988347315,"speed":90.58228710164383},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"jez-large-1.0.0","name":"Jez-large-1.0.0 (pragmaticfrontier, Gemma-4-E4B label-logit decision model, author server, self-hosted)","board":"open","source_url":"https://huggingface.co/pragmaticfrontier/jez-large-1.0.0","axes":{"intelligence":25.626746881681953,"calibration":83.15885665129233,"speed":90.75287985569767,"cost":83.31206101384694},"capability":{"score":54.39280176648714,"rank":204,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":170,"api_board_rank":null},"composite_score":14.186361224326426,"rank":237,"open_board_rank":190,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003599686115098442,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 712,537 input + 0 output tokens over 1,485 decisions = 479.822895623 tokens/decision. GPU-time cost USD 0.00359968611509844/1,000; old list-price cost USD 0.00959645791245791/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00359968611509844/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05969935749999422,"p50_s_adjusted":0.26939871499998846,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05969935749999422,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0810581660500361,"p95_s_adjusted":0.31211633210007217,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jez-small-1.0.0","name":"Jez-small-1.0.0 (pragmaticfrontier, Qwen3.5-0.8B label-logit decision model, author server, self-hosted)","board":"open","source_url":"https://huggingface.co/pragmaticfrontier/jez-small-1.0.0","axes":{"intelligence":4.560234849177464,"calibration":81.90437649433692,"speed":92.50829073456676,"cost":86.2098204316998},"capability":{"score":43.23230567175719,"rank":239,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":202,"api_board_rank":null},"composite_score":0.1310454200964957,"rank":294,"open_board_rank":241,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00288185849500294,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 707,876 input + 0 output tokens over 1,485 decisions = 476.684175084 tokens/decision. GPU-time cost USD 0.00288185849500294/1,000; old list-price cost USD 0.00476684175084175/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00288185849500294/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04254042100001243,"p50_s_adjusted":0.23508084200002485,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04254042100001243,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04437783359992409,"p95_s_adjusted":0.23875566719984817,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jiwo-0.8b","name":"jiwo 0.8b (Qwen3.5-0.8B backbone with a trained decision head, one forward pass, native probabilities, self-hosted)","board":"open","source_url":"https://huggingface.co/eljiwo/jiwo-0.8b","axes":{"intelligence":15.872104931583502,"calibration":78.54210067220151,"speed":94.31864739858605,"cost":84.96035415353933},"capability":{"score":47.20710280189251,"rank":229,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":193,"api_board_rank":null},"composite_score":4.108504374494453,"rank":262,"open_board_rank":213,"api_board_rank":null,"measured_in":"v1.6.9","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0031719149087301593,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 786,993 input + 0 output tokens over 1,500 decisions = 524.662 tokens/decision. GPU-time cost USD 0.00317191490873016/1,000; old list-price cost USD 0.00524662/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00317191490873016/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.019949944515246898,"p50_s_adjusted":0.1898998890304938,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.019949944515246898,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.022404843638651074,"p95_s_adjusted":0.19480968727730213,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":4.067767731552672,"capability":47.20710280189251,"axes":{"calibration":78.54210067220151,"cost":78.40361168685378,"intelligence":15.872104931583502,"speed":94.33479301471377},"rank":185,"open_board_rank":145,"api_board_rank":null}},{"key":"jiwo-4b","name":"jiwo 4b (Qwen3.5-4B backbone with a trained decision head, one forward pass, native probabilities, self-hosted)","board":"open","source_url":"https://huggingface.co/eljiwo/jiwo-4b","axes":{"intelligence":34.784391844693545,"calibration":85.66971124396188,"speed":92.17972487720941,"cost":78.63190282478453},"capability":{"score":60.227051544327715,"rank":169,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":137,"api_board_rank":null},"composite_score":30.25493665442467,"rank":192,"open_board_rank":152,"api_board_rank":null,"measured_in":"v1.6.9","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005155489490012448,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 786,993 input + 0 output tokens over 1,500 decisions = 524.662 tokens/decision. GPU-time cost USD 0.00515548949001245/1,000; old list-price cost USD 0.01573986/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00515548949001245/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0363162939902395,"p50_s_adjusted":0.222632587980479,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0363162939902395,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06095925677800551,"p95_s_adjusted":0.27191851355601104,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":28.962932355698086,"capability":60.227051544327715,"axes":{"calibration":85.66971124396188,"cost":64.0899740452639,"intelligence":34.784391844693545,"speed":92.44547609658223},"rank":103,"open_board_rank":71,"api_board_rank":null}},{"key":"joey-1","name":"baselabs joey-1 (2B)","board":"open","source_url":"https://huggingface.co/baselabs/joey-1","axes":{"intelligence":23.488157046202485,"calibration":87.3438060403445,"speed":92.05436718949184,"cost":89.05819789393655},"capability":{"score":55.415981543273496,"rank":198,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":164,"api_board_rank":null},"composite_score":11.596989171303864,"rank":244,"open_board_rank":197,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0023159369769846605,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 568,911 input + 1,489 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00231593697698466/1,000; old list-price estimate USD 0.00774151779717931/1,000 (deepinfra:Qwen/Qwen3.5-2B at USD 0.02/M input and USD 0.1/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00231593697698466/1,000. Estimated, not a bill. USD 0.00231593698 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00230963809 over all 1,500 items)."},"latency":{"p50_s_raw":0.04631088000002137,"p50_s_adjusted":0.24262176000004274,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.04631088000002137,"latency_basis":"api-end-to-end","p95_s_raw":0.05341138689997535,"p95_s_adjusted":0.25682277379995067,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"jpt-35b-a3b","name":"JPT-35B-A3B (Qwen3.5-35B-A3B fine-tune)","board":"open","source_url":"https://huggingface.co/kirp/jpt-35b-a3b","axes":{"intelligence":51.42208418374838,"calibration":89.9313809717458,"speed":90.38067871211373,"cost":72.4983280625774},"capability":{"score":70.67673257774709,"rank":98,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":74,"api_board_rank":null},"composite_score":72.1706145001871,"rank":74,"open_board_rank":57,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008255101128320605,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 764,897 input + 1,477 output tokens over 1,477 decisions = 518.872037915 tokens/decision. GPU-time cost USD 0.0082551011283206/1,000; old list-price cost USD 0.163085011848341/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0082551011283206/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07176362050813623,"p50_s_adjusted":0.2935272410162725,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07176362050813623,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08104638804506976,"p95_s_adjusted":0.3120927760901395,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":25.357657195103958,"capability":70.67673257774709,"axes":{"calibration":89.9313809717458,"cost":33.627578514713974,"intelligence":51.42208418374838,"speed":90.51950606084544},"rank":111,"open_board_rank":78,"api_board_rank":null}},{"key":"jpt-4b","name":"JPT-4B (kirp, Qwen3.5-4B LoRA)","board":"open","source_url":"https://huggingface.co/kirp/jpt-4b","axes":{"intelligence":40.7504160092812,"calibration":86.88996927941005,"speed":87.81344131912249,"cost":78.85487162848112},"capability":{"score":63.820192644345624,"rank":141,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":111,"api_board_rank":null},"composite_score":44.19584526267103,"rank":158,"open_board_rank":125,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005068011741545535,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 764,418 input + 1,485 output tokens over 1,485 decisions = 515.75959596 tokens/decision. GPU-time cost USD 0.00506801174154553/1,000; old list-price cost USD 0.0155927878787879/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00506801174154553/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06583588174544275,"p50_s_adjusted":0.2816717634908855,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06583588174544275,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.21868553704116492,"p95_s_adjusted":0.5873710740823298,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":42.17979190383397,"capability":63.820192644345624,"axes":{"calibration":86.88996927941005,"cost":64.21228687378368,"intelligence":40.7504160092812,"speed":87.95549105905897},"rank":76,"open_board_rank":49,"api_board_rank":null}},{"key":"jqv","name":"jqv (Qwen3-32B zero-shot)","board":"open","source_url":"https://github.com/Octalab-Inc/jqv","axes":{"intelligence":33.85815624201783,"calibration":89.87138320201578,"speed":82.47453178010326,"cost":54.72399117000985},"capability":{"score":61.8647697220168,"rank":156,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":124,"api_board_rank":null},"composite_score":25.8118135489931,"rank":201,"open_board_rank":160,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.032299834296411846,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 32B active parameters; throughput 9925.03125 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 861,145 input + 0 output tokens over 1,624 decisions = 530.261699507 tokens/decision. GPU-time cost USD 0.0322998342964118/1,000; old list-price cost USD 0.0424209359605911/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0322998342964118/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11767431150656193,"p50_s_adjusted":0.3853486230131239,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11767431150656193,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.6589441528078168,"p95_s_adjusted":1.4678883056156335,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":25.391439220985372,"capability":61.8647697220168,"axes":{"calibration":89.87138320201578,"cost":51.172592607797476,"intelligence":33.85815624201783,"speed":83.1032423977654},"rank":110,"open_board_rank":77,"api_board_rank":null}},{"key":"jul-fast","name":"jul fast (usejul/jul-decision-minicpm5-2b: MiniCPM5-2B fine-tune with a pointer head, option probabilities from one pass, self-hosted via `jul serve`)","board":"open","source_url":"https://huggingface.co/usejul/jul-decision-minicpm5-2b","axes":{"intelligence":7.4645122320968404,"calibration":77.79406843975872,"speed":90.5592978634667,"cost":68.9870979950853},"capability":{"score":42.62929033592778,"rank":245,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":208,"api_board_rank":null},"composite_score":0.5172342787655929,"rank":278,"open_board_rank":226,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.010808449559918753,"basis":"ESTIMATE (size-class reference; self-hosted open weights, no bookable listing for this base). openbmb/MiniCPM5-2B is an EXPLICIT entry in the frozen module's UNLISTED_BASE_MODELS table, so pricing_v15.reference() returns None without raising: the decision that there is no bookable listing for this base is already taken, and the only MiniCPM key in either dated snapshot is deepinfra:openbmb/MiniCPM-Llama3-V-2_5 at (0.34, 0.34), a VISION model and not a comparable reference. The row therefore carries its own labelled size-class estimate: deepinfra:Qwen/Qwen3.5-2B at USD 0.02 per 1M input and USD 0.10 per 1M output in the frozen 25 Sep snapshot - a 2B dense instruct model, the same size class as this 2B base - x this row's own measured tokens. Estimated, not charged. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone and any other decision is reversible by arithmetic from the raw. PRICING REFERENCE NOTE: if a bookable MiniCPM5-2B text listing has appeared since 25 Sep, this row reprices by one multiplication from the raw. receipts/COST-ALTERNATIVES-R58.json has the arithmetic for the alternatives already done."},"latency":{"p50_s_raw":0.04378877401177306,"p50_s_adjusted":0.2375775480235461,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04378877401177306,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11002687854197575,"p95_s_adjusted":0.3700537570839515,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":4,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.5174168972906218,"capability":42.76419280473996,"axes":{"calibration":78.06387337738309,"cost":68.9870979950853,"intelligence":7.4645122320968404,"speed":90.69403874154062},"rank":204,"open_board_rank":161,"api_board_rank":null}},{"key":"julia-1","name":"Julia-1 (Supersonic Labs)","board":"open","source_url":"https://huggingface.co/SupersonicLabs/Julia-1","axes":{"intelligence":2.4412732607743255,"calibration":17.26803648238681,"speed":77.74663940674506,"cost":89.74859303626371},"capability":{"score":9.854654871580568,"rank":280,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":241,"api_board_rank":null},"composite_score":0.01939975687732831,"rank":303,"open_board_rank":250,"api_board_rank":null,"measured_in":"v1.6.15","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0021964107744107744,"basis":"ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff for these weights. The frozen pricing_v15 path is asserted with ONE authorised, scoped correction: the frozen pricing update added the declared base jhu-clsp/mmBERT-small to pricing_v15.UNLISTED_BASE_MODELS, exactly as openbmb/MiniCPM5-2B and answerdotai/ModernBERT-base already are, so reference() returns no market reference for it rather than raising and the ancestry can be carried in base_model instead of being dropped to null. The addition changes no other row: it was proved behaviour-identical over every registry entry on the box before it was installed (r108/receipts/PRICING-EQUIV-R108.json) and independently reviewed (r108/receipts/REVIEW-PRICING-R108.json). The published 24 Sep estimate rule then gives a clearly labelled size-class estimate: the base-size-encoder line the PUBLISHED v1.6 field already carries - USD 0.005 per 1M input and 0.0 per 1M output, the line modernbert-jev-v2, tacet-sonata and watt-flash-0.1 carry there (no row of THIS draw carries it, and lev-350m carries 0.01/0.0, not this line) (four base-size encoders in the dated 25 Sep 2026 DeepInfra snapshot sha256 84eb7926... all price at exactly that) - so the frozen pricing reference is unchanged. x this row's own measured tokens: usage.input_tokens comes from the author's OWN audit function engine.encoding_info() and usage.output_tokens is 0 by construction (one encoder forward pass read at marker positions, nothing generated). Estimated, not charged. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.002196 per 1,000 decisions (0.034x the official cost cap USD 0.064595), against USD 0.002174 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.3325943410018226,"p50_s_adjusted":0.8151886820036452,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.3325943410018226,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.9554998249011987,"p95_s_adjusted":2.0609996498023975,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.019410466159695165,"capability":9.854654871580568,"axes":{"calibration":17.26803648238681,"cost":89.74859303626371,"intelligence":2.4412732607743255,"speed":79.42119248319958},"rank":230,"open_board_rank":186,"api_board_rank":null}},{"key":"kahn1","name":"Kahn1 4B (Okura66, Qwen3.5-4B LoRA merge, author's sysone engine)","board":"open","source_url":"https://huggingface.co/Okura66/Kahn1-Qwen3.5-4B","axes":{"intelligence":39.71971801341685,"calibration":83.76747077065221,"speed":89.23426082115222,"cost":64.66738128007644},"capability":{"score":61.74359439203453,"rank":158,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":126,"api_board_rank":null},"composite_score":39.57470163141661,"rank":175,"open_board_rank":138,"api_board_rank":null,"measured_in":"v1.6.1","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-07","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.015057537635197281,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,488,569 input + 0 output tokens over 1,624 decisions = 1532.37007389 tokens/decision. GPU-time cost USD 0.0150575376351973/1,000; old list-price cost USD 0.0459711022167488/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0150575376351973/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07760014399536885,"p50_s_adjusted":0.30520028799073773,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07760014399536885,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12041548872366546,"p95_s_adjusted":0.3908309774473309,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":37.0242246815327,"capability":61.74359439203453,"axes":{"calibration":83.76747077065221,"cost":50.12545250055582,"intelligence":39.71971801341685,"speed":89.97531755002697},"rank":89,"open_board_rank":59,"api_board_rank":null}},{"key":"karpo_decision_31b","name":"Karpo-Decision-31B","board":"open","source_url":"https://huggingface.co/MachinePulseAI/karpo-decision-31B","axes":{"intelligence":80.1759451541962,"calibration":94.14306241688593,"speed":88.29089098203782,"cost":57.9400170643567},"capability":{"score":87.15950378554106,"rank":3,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":3,"api_board_rank":null},"composite_score":77.39919232893257,"rank":12,"open_board_rank":7,"api_board_rank":null,"measured_in":"v1.8.9","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.9","measured_on":"2026-11-10","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.025234774673846376,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.0252347746738464/1,000; old list-price estimate USD 0.0379665815983882/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0252347746738464/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.0821957040000143,"p50_s_adjusted":0.3143914080000286,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0821957040000143,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16072749300005854,"p95_s_adjusted":0.47145498600011704,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.9","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"kettei-chub","name":"kettei-chub (Sqwish)","board":"api","source_url":"https://docs.sqwish.ai","axes":{"intelligence":62.182714942969874,"calibration":84.10309591585386,"speed":94.89692360127322,"cost":62.69047197239138},"capability":{"score":73.14290542941187,"rank":78,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":22},"composite_score":73.44712821007478,"rank":52,"open_board_rank":null,"api_board_rank":15,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.017524713804713803,"basis":"Sqwish public price list, USD 0.04/M input, output not billed (docs.sqwish.ai/guides/pricing, fetched 16 Oct 2026 16:17Z, saved copy sha256 698a29e5...): kettei chub at USD 0.04 per million input tokens for decisions; \"The models generate no output tokens, and none are billed\"; \"standard list prices, the same since launch on 14 October 2026\". Counted as the manufacturer's standard launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap, d1-ultra, Sage 1.4 and Nimble), through a scoped re-freeze of the frozen pricing module that adds the two kettei keys and nothing else, with a zero-difference proof for every other registry entry. The once-per-inbox USD 5.00 starter credit is a trial credit and is not used. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (650,605 input tokens over the 1,485 answered items, 0 output tokens), divided by the 1,485 items of the common cost basis the scorer uses; the 15 items refused with HTTP 422 count as zero tokens and as wrong answers, as every refused request on this draw does (d1-ultra's identical 15 refusals included). The unchanged runner also wrote a conservative cost ceiling into those 15 refused rows (65,536 input tokens each, the model's input limit); the scoring copy drops that one field on those 15 rows only. With the ceiling charged on all 1,500 items (the former all-items basis), the figure would be USD 0.043564 per 1,000 decisions (0.674x the official cost cap); that is a sensitivity, not this row's cost, and on the common basis the 15 refused items are not part of cost at all. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.017525 per 1,000 decisions (0.271x the official cost cap USD 0.064595), against USD 0.017349 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.16808052101987414,"p50_s_adjusted":0.16808052101987414,"adjustment":"none (API)","p50_s_measured":0.16808052101987414,"latency_basis":"api-end-to-end","p95_s_raw":0.19265942253987306,"p95_s_adjusted":0.19265942253987306,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":73.47624549745721,"capability":73.14290542941187,"axes":{"calibration":84.10309591585386,"cost":62.69047197239138,"intelligence":62.182714942969874,"speed":95.09167642607092},"rank":11,"open_board_rank":null,"api_board_rank":9}},{"key":"kettei-fat","name":"kettei-fat (Sqwish)","board":"api","source_url":"https://docs.sqwish.ai","axes":{"intelligence":79.83918785475791,"calibration":92.78078361709031,"speed":92.90799647527135,"cost":57.40773420072094},"capability":{"score":86.3099857359241,"rank":4,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":1},"composite_score":77.69122909321317,"rank":9,"open_board_rank":null,"api_board_rank":5,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0262870707070707,"basis":"Sqwish public price list, USD 0.06/M input, output not billed (docs.sqwish.ai/guides/pricing, fetched 16 Oct 2026 16:17Z, saved copy sha256 698a29e5...): kettei fat at USD 0.06 per million input tokens for decisions; \"The models generate no output tokens, and none are billed\"; \"standard list prices, the same since launch on 14 October 2026\". Counted as the manufacturer's standard launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap, d1-ultra, Sage 1.4 and Nimble), through a scoped re-freeze of the frozen pricing module that adds the two kettei keys and nothing else, with a zero-difference proof for every other registry entry. The once-per-inbox USD 5.00 starter credit is a trial credit and is not used. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (650,605 input tokens over the 1,485 answered items, 0 output tokens), divided by the 1,485 items of the common cost basis the scorer uses; the 15 items refused with HTTP 422 count as zero tokens and as wrong answers, as every refused request on this draw does (d1-ultra's identical 15 refusals included). The unchanged runner also wrote a conservative cost ceiling into those 15 refused rows (65,536 input tokens each, the model's input limit); the scoring copy drops that one field on those 15 rows only. With the ceiling charged on all 1,500 items (the former all-items basis), the figure would be USD 0.065346 per 1,000 decisions (1.012x the official cost cap); that is a sensitivity, not this row's cost, and on the common basis the 15 refused items are not part of cost at all. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.026287 per 1,000 decisions (0.407x the official cost cap USD 0.064595), against USD 0.026024 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.20237550800084136,"p50_s_adjusted":0.20237550800084136,"adjustment":"none (API)","p50_s_measured":0.20237550800084136,"latency_basis":"api-end-to-end","p95_s_raw":0.2529544946068199,"p95_s_adjusted":0.2529544946068199,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":77.74062514641476,"capability":86.3099857359241,"axes":{"calibration":92.78078361709031,"cost":57.40773420072094,"intelligence":79.83918785475791,"speed":93.19124027169539},"rank":1,"open_board_rank":null,"api_board_rank":1}},{"key":"kev-0.5b","name":"kev 0.5B","board":"open","source_url":"https://github.com/jaredpalmer/kev","axes":{"intelligence":4.598433868646836,"calibration":61.20618525995795,"speed":92.39061085463217,"cost":84.19572866753369},"capability":{"score":32.90230956430239,"rank":264,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":226,"api_board_rank":null},"composite_score":0.13190015086675236,"rank":293,"open_board_rank":240,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0033636364062866533,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.5B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,376 input + 155,176 output tokens over 1,624 decisions = 556.374384236 tokens/decision. GPU-time cost USD 0.00336363640628665/1,000; old list-price cost USD 0.00460822660098522/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00336363640628665/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0440487046726048,"p50_s_adjusted":0.2380974093452096,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0440487046726048,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04610281793866307,"p95_s_adjusted":0.24220563587732613,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.131588623265787,"capability":32.90230956430239,"axes":{"calibration":61.20618525995795,"cost":80.09398520305423,"intelligence":4.598433868646836,"speed":92.39904218127765},"rank":220,"open_board_rank":176,"api_board_rank":null}},{"key":"kev-0.6b","name":"kev 0.6B (research preview)","board":"open","source_url":"https://github.com/jaredpalmer/kev","axes":{"intelligence":7.598789169415934,"calibration":66.26655558235618,"speed":91.48734752715623,"cost":84.20067552024135},"capability":{"score":36.93267237588606,"rank":257,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":220,"api_board_rank":null},"composite_score":0.5450619596118789,"rank":277,"open_board_rank":225,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003362359526497381,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.6B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,375 input + 154,834 output tokens over 1,624 decisions = 556.16317734 tokens/decision. GPU-time cost USD 0.00336235952649738/1,000; old list-price cost USD 0.00460822044334975/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00336235952649738/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.056589895160868764,"p50_s_adjusted":0.26317979032173755,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.056589895160868764,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.059890919481404126,"p95_s_adjusted":0.2697818389628083,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.5431213026023877,"capability":36.93267237588606,"axes":{"calibration":66.26655558235618,"cost":80.09400261254245,"intelligence":7.598789169415934,"speed":91.51497362122684},"rank":203,"open_board_rank":160,"api_board_rank":null}},{"key":"kev-27b","name":"Kev 27B (Kev 1.0, Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/jaredpalmer/kev-27b","axes":{"intelligence":64.16071886717361,"calibration":93.71720359195915,"speed":82.8278891625431,"cost":56.1644780115139},"capability":{"score":78.93896122956639,"rank":42,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":32,"api_board_rank":null},"composite_score":71.25763456915091,"rank":85,"open_board_rank":67,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.028919052939090185,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 663,658 input + 167,418 output tokens over 1,477 decisions = 562.678402167 tokens/decision. GPU-time cost USD 0.0289190529390902/1,000; old list-price cost USD 0.528768016249154/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0289190529390902/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.18723015300929546,"p50_s_adjusted":0.524460306018591,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.18723015300929546,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.4221282738435546,"p95_s_adjusted":0.9942565476871092,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":5.77223551264444,"capability":78.93896122956639,"axes":{"calibration":93.71720359195915,"cost":18.302044660887503,"intelligence":64.16071886717361,"speed":83.65684843680415},"rank":184,"open_board_rank":144,"api_board_rank":null}},{"key":"kev-4b","name":"kev 4B (research preview)","board":"open","source_url":"https://github.com/jaredpalmer/kev","axes":{"intelligence":19.510707720813777,"calibration":68.45285936403559,"speed":90.22194613896767,"cost":77.89103360168605},"capability":{"score":43.98178354242468,"rank":236,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":199,"api_board_rank":null},"composite_score":6.78365205932627,"rank":257,"open_board_rank":208,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005457145668217454,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,375 input + 153,531 output tokens over 1,624 decisions = 555.360837438 tokens/decision. GPU-time cost USD 0.00545714566821745/1,000; old list-price cost USD 0.0138246613300493/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00545714566821745/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0722477997187525,"p50_s_adjusted":0.29449559943750503,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0722477997187525,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08632310815155504,"p95_s_adjusted":0.3226462163031101,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":6.611791819903415,"capability":43.98178354242468,"axes":{"calibration":68.45285936403559,"cost":65.78036497095258,"intelligence":19.510707720813777,"speed":90.46585183268016},"rank":182,"open_board_rank":142,"api_board_rank":null}},{"key":"kev-8b","name":"kev 8B (research preview)","board":"open","source_url":"https://github.com/jaredpalmer/kev","axes":{"intelligence":29.26217941280341,"calibration":70.7561729087212,"speed":88.45850101024836,"cost":68.86568213459405},"capability":{"score":50.0091761607623,"rank":218,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":182,"api_board_rank":null},"composite_score":18.48094036640983,"rank":218,"open_board_rank":175,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01090964441328195,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 8B active parameters; throughput 30762.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,375 input + 153,147 output tokens over 1,624 decisions = 555.124384236 tokens/decision. GPU-time cost USD 0.0109096444132819/1,000; old list-price cost USD 0.0968237438423645/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0109096444132819/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07640513079240918,"p50_s_adjusted":0.3028102615848184,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07640513079240918,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16047744499519467,"p95_s_adjusted":0.47095488999038937,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":10.62283914209564,"capability":50.0091761607623,"axes":{"calibration":70.7561729087212,"cost":40.42054386073457,"intelligence":29.26217941280341,"speed":88.97143473610706},"rank":169,"open_board_rank":130,"api_board_rank":null}},{"key":"kev-9b-v2","name":"Kev-9B v2 (Jared Palmer, Qwen3.5-9B-Base LoRA + pointer head)","board":"open","source_url":"https://huggingface.co/jaredpalmer/kev-9b","axes":{"intelligence":36.22478886598313,"calibration":86.77132815303852,"speed":89.01765799382792,"cost":null},"capability":{"score":61.49805850951083,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base and raises rather than estimating; one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 3948e3f1..., and the later re-freezes, which only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base (the module RAISES 'base model lacks an explicit market-listing decision'); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens as reported by the author's server, kept as sent (input 662,682; output 164,737 is the author's billing-style count of the serialised answers, not generated text)."},"latency":{"p50_s_raw":0.05437049572356045,"p50_s_adjusted":0.2587409914471209,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05437049572356045,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16729193984065202,"p95_s_adjusted":0.484583879681304,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":61.49805850951083,"axes":{"calibration":86.77132815303852,"cost":null,"intelligence":36.22478886598313,"speed":89.1575041920417},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"knowledgator_gliclass_v4_9b","name":"GLiClass v4 9B (Knowledgator)","board":"open","source_url":"https://huggingface.co/knowledgator/gliclass-v4-9b","axes":{"intelligence":39.63748168273355,"calibration":86.76027774140783,"speed":85.63164752763427,"cost":70.90886953813762},"capability":{"score":63.19887971207069,"rank":146,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":115,"api_board_rank":null},"composite_score":40.19829244906556,"rank":171,"open_board_rank":135,"api_board_rank":null,"measured_in":"v1.7.9","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.9","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009326191927400784,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 626,409 input + 0 output tokens over 1,485 decisions = 421.824242424 tokens/decision. GPU-time cost USD 0.00932619192740078/1,000; old list-price cost USD 0.0421824242424242/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00932619192740078/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1347206760000006,"p50_s_adjusted":0.4194413520000012,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1347206760000006,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2509372574999816,"p95_s_adjusted":0.6518745149999632,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.9","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"knowline-26b-a4b-gen2.5","name":"KnowLine 26B-A4B-Gen2.5","board":"open","source_url":"https://huggingface.co/PelaAI/KnowLine-26B-A4B-Gen2.5","axes":{"intelligence":65.23825939945094,"calibration":89.44042524156703,"speed":89.461278844294,"cost":62.33805091041278},"capability":{"score":77.33934232050899,"rank":53,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":39,"api_board_rank":null},"composite_score":74.44818685146753,"rank":37,"open_board_rank":25,"api_board_rank":null,"measured_in":"v1.7.18","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.18","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.018005214641140434,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,531,517 input + 1,485 output tokens over 1,485 decisions = 1032.32457912 tokens/decision. GPU-time cost USD 0.0180052146411404/1,000; old list-price cost USD 0.0931192121212121/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0180052146411404/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07892856999990272,"p50_s_adjusted":0.30785713999980546,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07892856999990272,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10886238505004257,"p95_s_adjusted":0.36772477010008514,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.18","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"knowline-31b-gen1","name":"KnowLine 31B-Gen1","board":"open","source_url":"https://huggingface.co/PelaAI/KnowLine-31B-Gen1","axes":{"intelligence":77.43747194470296,"calibration":92.98946060661117,"speed":81.4421691177115,"cost":46.490946925756504},"capability":{"score":85.21346627565707,"rank":7,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":5,"api_board_rank":null},"composite_score":60.18983325088006,"rank":129,"open_board_rank":99,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.06076242562235659,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10271.2448706 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 1,531,517 input + 1,485 output tokens over 1,485 decisions = 1032.32457912 tokens/decision. GPU-time cost USD 0.0607624256223566/1,000; old list-price cost USD 0.0931592121212121/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0607624256223566/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.2546755749999079,"p50_s_adjusted":0.6593511499998158,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2546755749999079,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.4690468806501485,"p95_s_adjusted":1.088093761300297,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"knowline-35b-a3b-gen1.5","name":"KnowLine 35B-A3B-Gen1.5","board":"open","source_url":"https://huggingface.co/PelaAI/KnowLine-35B-A3B-Gen1.5","axes":{"intelligence":60.31798461023689,"calibration":92.98322836048867,"speed":91.72458132132627,"cost":63.558453547882905},"capability":{"score":76.65060648536277,"rank":59,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":44,"api_board_rank":null},"composite_score":74.11637231061094,"rank":40,"open_board_rank":27,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0163952564839634,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,528,836 input + 1,485 output tokens over 1,485 decisions = 1030.51919192 tokens/decision. GPU-time cost USD 0.0163952564839634/1,000; old list-price cost USD 0.322974747474747/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0163952564839634/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04758722399998305,"p50_s_adjusted":0.2451744479999661,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04758722399998305,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06209975099991425,"p95_s_adjusted":0.2741995019998285,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"knowline-4b-gen5.5","name":"KnowLine 4B-Gen5.5","board":"open","source_url":"https://huggingface.co/PelaAI/KnowLine-4B-Gen5.5","axes":{"intelligence":46.410576498288506,"calibration":85.97018847365896,"speed":92.9847804332849,"cost":69.83660860914662},"capability":{"score":66.19038248597374,"rank":126,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":99,"api_board_rank":null},"composite_score":59.16168733982439,"rank":132,"open_board_rank":102,"api_board_rank":null,"measured_in":"v1.7.18","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.18","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.010126197176840546,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,528,836 input + 1,485 output tokens over 1,485 decisions = 1030.51919192 tokens/decision. GPU-time cost USD 0.0101261971768405/1,000; old list-price cost USD 0.0310355757575758/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0101261971768405/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02987294400000451,"p50_s_adjusted":0.209745888000009,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02987294400000451,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.044894291500048665,"p95_s_adjusted":0.2397885830000973,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.18","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"krino-2b","name":"Krino-2B","board":"open","source_url":"https://huggingface.co/Guru0381/krino-2b","axes":{"intelligence":21.99889248592056,"calibration":62.82741496733333,"speed":90.96816394625492,"cost":84.1923853867817},"capability":{"score":42.413153726626945,"rank":246,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":209,"api_board_rank":null},"composite_score":9.191434964036928,"rank":248,"open_board_rank":201,"api_board_rank":null,"measured_in":"v1.6.10","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0033644996472663143,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 662,682 input + 163,746 output tokens over 1,485 decisions = 556.517171717 tokens/decision. GPU-time cost USD 0.00336449964726631/1,000; old list-price cost USD 0.0199516767676768/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00336449964726631/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.031744987005367875,"p50_s_adjusted":0.21348997401073574,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.031744987005367875,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11240282154467421,"p95_s_adjusted":0.3748056430893484,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":8.734255211215622,"capability":42.413153726626945,"axes":{"calibration":62.82741496733333,"cost":61.00061799127391,"intelligence":21.99889248592056,"speed":91.85485240061512},"rank":176,"open_board_rank":137,"api_board_rank":null}},{"key":"kushal-gemma4-31b-it-autoloops","name":"Autoloops – Gemma 4 31B IT","board":"api","source_url":"https://benchmarkheaven.com/jev-models/kushal-gemma4-31b-it-autoloops","axes":{"intelligence":58.1,"calibration":87.31,"speed":84.44,"cost":40.47},"capability":{"score":72.705,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 3.0× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":40.17,"rank":172,"open_board_rank":null,"api_board_rank":37,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.09648663282571913,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.2/M input, 0.45/M output (spec kushal-gemma4-31b-it-autoloops.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.584,"p50_s_adjusted":0.584,"adjustment":"none (API)","p50_s_measured":0.584,"latency_basis":"api-end-to-end","p95_s_raw":0.616,"p95_s_adjusted":0.616,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":40.18,"capability":72.7,"axes":{"calibration":87.31,"cost":40.47,"intelligence":58.1,"speed":84.51},"rank":82,"open_board_rank":null,"api_board_rank":29}},{"key":"launchables-myles-g-12b","name":"Myles-G 12B (Launchables)","board":"api","source_url":"https://www.launchables.ai/licensing","axes":{"intelligence":57.25409127121432,"calibration":84.04872871623093,"speed":85.292646608578,"cost":99.35683144033557},"capability":{"score":70.65140999372262,"rank":99,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":24},"composite_score":78.19688312813278,"rank":5,"open_board_rank":null,"api_board_rank":3,"measured_in":"v1.6.38","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0010506037629629628,"basis":"Public launch list price, counted under interpretation I-1 (a manufacturer's standard, non-promotional list price counts from day 1): GBP 0.0015 per million input tokens, output free, from launchables.ai/docs/deploy/hosted fetched 26 Oct 2026 17:38 UTC. USD 0.0020020 per million input tokens from the ECB euro reference rates of 20 Oct 2026 (GBP 1 = USD 1.33468); prices exclude VAT. Provider-reported usage: 779,294 input / 1,485 output tokens on the common 1,485-item basis; 989,607 input / 1,500 output over all 1,500 items. Common basis USD 0.00105060 per 1,000 decisions; all-item basis USD 0.00132080. The tariff applied to the row's own provider-reported tokens; not a bill and not a quote. Saved copy of the vendor page sha256 956fa709. Price corrected in v1.7.9; earlier releases showed a same-base ESTIMATE of USD 0.0264 per 1,000."},"latency":{"p50_s_raw":0.4840246535022743,"p50_s_adjusted":0.4840246535022743,"adjustment":"none (API)","p50_s_measured":0.4840246535022743,"latency_basis":"api-end-to-end","p95_s_raw":0.6107561607146635,"p95_s_adjusted":0.6107561607146635,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.40656153142275,"capability":70.65140999372262,"axes":{"calibration":84.04872871623093,"cost":57.35738348807602,"intelligence":57.25409127121432,"speed":85.64932567577961},"rank":34,"open_board_rank":null,"api_board_rank":18}},{"key":"launchables-myles-g-31b","name":"Myles-G 31B (Launchables)","board":"api","source_url":"https://www.launchables.ai/licensing","axes":{"intelligence":73.64186833978687,"calibration":83.99531596546265,"speed":85.94853930142175,"cost":51.11161516579885},"capability":{"score":78.81859215262476,"rank":44,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":11},"composite_score":70.56596417573436,"rank":91,"open_board_rank":null,"api_board_rank":21,"measured_in":"v1.6.36","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0426199393939394,"basis":"ESTIMATE, v1.5-M2 base-model reference rule. Launchables publishes a list price for Myles-G 31B (launchables.ai/licensing, fetched 21 Oct 2026, saved copy sha256 ad4736d5...): GBP 0.005 per million input tokens, output free; that is USD 0.0066734 per million input tokens at the ECB reference rate of 20 Oct 2026 (GBP 1 = USD 1.33468), prices exclude VAT. Under the method, the base-model reference price applies as an ESTIMATE for this vendor-hosted row, as for attest-31b-v1.0: base model google/gemma-4-31B-it through the frozen pricing_v15 module (sha256 1f4d1f7b1ce97ab27b938043444b9c2757e4a9dbecc799a803122588401f5bf1) reference() -> google/gemma-4-31b-it on OpenRouter, USD 0.09 per million input tokens, USD 0.34 per million output tokens. The vendor list price is shown for information and is not used for the Cost axis. Applied to this row's own provider-reported tokens (907,107 input and 1,500 output tokens over all 1,500 items; 697,619 input and 1,485 output tokens over the 1,485 items of the common cost basis; the 15 truncated items count for no row's cost). Estimated, not a vendor bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them in full)."},"latency":{"p50_s_raw":0.4449399680015631,"p50_s_adjusted":0.4449399680015631,"adjustment":"none (API)","p50_s_measured":0.4449399680015631,"latency_basis":"api-end-to-end","p95_s_raw":0.5712742523639462,"p95_s_adjusted":0.5712742523639462,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.6763821133842,"capability":78.81859215262476,"axes":{"calibration":83.99531596546265,"cost":51.11161516579885,"intelligence":73.64186833978687,"speed":86.60775159362743},"rank":24,"open_board_rank":null,"api_board_rank":15}},{"key":"laya","name":"Laya (Convai Innovations, ModernBERT-large 421M)","board":"open","source_url":"https://huggingface.co/convaiinnovations/laya","axes":{"intelligence":1.7910341507793501,"calibration":63.22326564159259,"speed":72.744382846694,"cost":84.89402784895724},"capability":{"score":32.50714989618597,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.008558710204581235,"rank":306,"open_board_rank":253,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0031881034482758625,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.9323667287826538,"p50_s_adjusted":2.0147334575653075,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.9323667287826538,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.244210566068068,"p95_s_adjusted":2.638421132136136,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.008560099345041717,"capability":32.50714989618597,"axes":{"calibration":63.22326564159259,"cost":84.89402784895724,"intelligence":1.7910341507793501,"speed":73.26302890816496},"rank":233,"open_board_rank":189,"api_board_rank":null}},{"key":"laya-multilingual","name":"Laya multilingual","board":"open","source_url":"https://huggingface.co/convaiinnovations/laya/tree/1c5edc17a7acd8701df6fc341c0d179f1c62c982/multilingual","axes":{"intelligence":0.2879184668737643,"calibration":33.655846666999985,"speed":76.95139513446136,"cost":82.1551651086044},"capability":{"score":16.971882566936873,"rank":278,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":239,"api_board_rank":null},"composite_score":0.00003759412198193081,"rank":308,"open_board_rank":255,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003933940886699508,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: documented hosted-model estimate; no exact base-model floor applies. $0.01/M input, $0/M output; same-class hosted-encoder estimate; no exact-base market floor."},"latency":{"p50_s_raw":0.4018409790005535,"p50_s_adjusted":0.953681958001107,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.4018409790005535,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.9828569027566117,"p95_s_adjusted":2.1157138055132236,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.000037595668430455566,"capability":16.971882566936873,"axes":{"calibration":33.655846666999985,"cost":82.1551651086044,"intelligence":0.2879184668737643,"speed":77.82044954853706},"rank":235,"open_board_rank":191,"api_board_rank":null}},{"key":"laya-typed-decisions","name":"Laya typed-decisions","board":"open","source_url":"https://huggingface.co/convaiinnovations/laya/tree/1c5edc17a7acd8701df6fc341c0d179f1c62c982/typed-decisions","axes":{"intelligence":5.471270906587037,"calibration":82.96996460403702,"speed":69.34079458737523,"cost":82.52942826196218},"capability":{"score":44.22061775531203,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.5× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.21636634154218096,"rank":286,"open_board_rank":233,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003822543103448276,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: documented hosted-model estimate; no exact base-model floor applies. $0.01/M input, $0/M output; same-class hosted-encoder estimate; no exact-base market floor."},"latency":{"p50_s_raw":1.0015642909929738,"p50_s_adjusted":2.1531285819859476,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":1.0015642909929738,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.6278415059503457,"p95_s_adjusted":5.405683011900692,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.21666109526429145,"capability":44.22061775531203,"axes":{"calibration":82.96996460403702,"cost":82.52942826196218,"intelligence":5.471270906587037,"speed":70.81965680870755},"rank":214,"open_board_rank":170,"api_board_rank":null}},{"key":"lev","name":"lev (Interfaze AI, Qwen3.5-4B + LoRA r32, label-token readout + candidate-path head, per-bucket temperatures)","board":"open","source_url":"https://huggingface.co/interfaze-ai/lev","axes":{"intelligence":41.084702071234844,"calibration":82.32371666823741,"speed":87.52560274254004,"cost":76.34084325677026},"capability":{"score":61.70420936973613,"rank":160,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":128,"api_board_rank":null},"composite_score":44.265822247960145,"rank":156,"open_board_rank":124,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006146651043081042,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,015,861 input + 0 output tokens over 1,624 decisions = 625.530172414 tokens/decision. GPU-time cost USD 0.00614665104308104/1,000; old list-price cost USD 0.0187659051724138/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00614665104308104/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10547674997360446,"p50_s_adjusted":0.36095349994720893,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10547674997360446,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16988290739536752,"p95_s_adjusted":0.489765814790735,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":42.177760514319104,"capability":61.70420936973613,"axes":{"calibration":82.32371666823741,"cost":61.79891447724963,"intelligence":41.084702071234844,"speed":88.00251499520694},"rank":77,"open_board_rank":50,"api_board_rank":null}},{"key":"lev-350m","name":"lev-350m (Franck Verrot, LFM2.5-350M)","board":"open","source_url":"https://github.com/franckverrot/lev","axes":{"intelligence":4.827562367090417,"calibration":81.86424479313416,"speed":93.7625770756901,"cost":80.03797760430007},"capability":{"score":43.34590358011229,"rank":238,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":201,"api_board_rank":null},"composite_score":0.1537555644865392,"rank":291,"open_board_rank":238,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00462807881773399,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.022485756999230944,"p50_s_adjusted":0.19497151399846188,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.022485756999230944,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.032830379250663094,"p95_s_adjusted":0.21566075850132618,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.15377137469809216,"capability":43.34590358011229,"axes":{"calibration":81.86424479313416,"cost":80.03797760430007,"intelligence":4.827562367090417,"speed":93.98230372494703},"rank":218,"open_board_rank":174,"api_board_rank":null}},{"key":"liquid-d1","name":"d1 (Liquid AI)","board":"api","source_url":"https://www.liquid.ai/blog/d1-decision-model","axes":{"intelligence":62.05738930369758,"calibration":86.78599512136702,"speed":87.8529201384932,"cost":62.84038569943876},"capability":{"score":74.4216922125323,"rank":67,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":17},"composite_score":72.81636405537589,"rank":64,"open_board_rank":null,"api_board_rank":16,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.017324224779959378,"basis":"Liquid AI d1 public API tariff USD 0.04 per 1M input tokens, no output tokens (Vercel AI Gateway listing https://vercel.com/ai-gateway/models/d1, read 6 Oct 2026; equals Liquid API usage.cost); measured on the v1.6.1 common cost basis from this run's token usage"},"latency":{"p50_s_raw":0.26902734860777855,"p50_s_adjusted":0.26902734860777855,"adjustment":"none (API)","p50_s_measured":0.26902734860777855,"latency_basis":"api-end-to-end","p95_s_raw":0.609412775933742,"p95_s_adjusted":0.609412775933742,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":73.0491258431614,"capability":74.4216922125323,"axes":{"calibration":86.78599512136702,"cost":62.84038569943876,"intelligence":62.05738930369758,"speed":89.22497122158413},"rank":13,"open_board_rank":null,"api_board_rank":10}},{"key":"liquid-d1-3b","name":"Liquid AI d1-3B","board":"open","source_url":"https://huggingface.co/LiquidAI/d1-3B","axes":{"intelligence":22.13919485880221,"calibration":76.87411794897184,"speed":94.0900145591257,"cost":50.56108621702079},"capability":{"score":49.50665640388703,"rank":219,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":183,"api_board_rank":null},"composite_score":8.853035001087052,"rank":249,"open_board_rank":202,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.044459420000000006,"basis":"ESTIMATE (*est.), size-class reference; self-hosted open weights, no bookable listing for this base. The declared base LiquidAI/LFM2.5-VL-3B is not a key of any frozen pricing table, and the frozen 25 Sep 2026 snapshot's only Liquid entry is liquid/lfm-2.5-2.6b:free at (0, 0) - a free route, which may not price a row (free-tier rule: use the published paid price or the estimate). Method, as aplomb-1, wald-4b-v2, jiwo-* and the ClassOne trio carry: the DeepInfra Qwen3.5 size-class reference from the frozen snapshot, rounded ONE CLASS UP so the estimate can only OVERSTATE this row's cost - deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output, for a 3B model. Applied to this row's own measured usage: 2,222,971 input tokens and 0 output tokens over 1,500 decisions. The 2B rate (0.02, 0.10) would give 0.029639613333333335 and was deliberately NOT used because it understates. Precedent for a LiquidAI base priced as a labelled estimate: the live lev-350m row (LiquidAI/LFM2.5-350M, in UNLISTED_BASE_MODELS, estimate 0.00462807881773399 per 1,000)."},"latency":{"p50_s_raw":0.016631776001304388,"p50_s_adjusted":0.18326355200260877,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.016631776001304388,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03138795199338346,"p95_s_adjusted":0.21277590398676693,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":8.855712627780655,"capability":49.50665640388703,"axes":{"calibration":76.87411794897184,"cost":50.56108621702079,"intelligence":22.13919485880221,"speed":94.32773191369529},"rank":175,"open_board_rank":136,"api_board_rank":null}},{"key":"liquid-d1-omni-600m","name":"Liquid AI d1-omni-600M","board":"open","source_url":"https://huggingface.co/LiquidAI/d1-omni-600M","axes":{"intelligence":3.2615852143588597,"calibration":62.32062493293682,"speed":93.99244100826726,"cost":73.1039158657426},"capability":{"score":32.79110507364784,"rank":265,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":227,"api_board_rank":null},"composite_score":0.04905613733213806,"rank":300,"open_board_rank":247,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00788018,"basis":"ESTIMATE (*est.), size-class reference; self-hosted open weights, no bookable listing for this base. The declared text base LiquidAI/LFM2.5-Encoder-350M is not a key of any frozen pricing table. Method as above: deepinfra:Qwen/Qwen3.5-0.8B at USD 0.01 per 1M input and USD 0.05 per 1M output - the class above this 600M model - applied to its own measured usage, 1,182,027 input tokens and 0 output tokens over 1,500 decisions. The frozen base-size ENCODER rate (0.005, 0.0), which the published tacet-sonata and lev-350m rows carry, would give 0.00394009 and was deliberately NOT used: it prices a text-only encoder, while d1-omni also runs SigLIP2 vision and FastConformer audio towers, so the class above is the conservative choice and keeps the two Liquid rows on one method."},"latency":{"p50_s_raw":0.02162825409322977,"p50_s_adjusted":0.19325650818645954,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02162825409322977,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.028179109469056128,"p95_s_adjusted":0.20635821893811224,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.049057448753865106,"capability":32.79110507364784,"axes":{"calibration":62.32062493293682,"cost":73.1039158657426,"intelligence":3.2615852143588597,"speed":94.0744546594334},"rank":226,"open_board_rank":182,"api_board_rank":null}},{"key":"liquid-d1-ultra","name":"d1-ultra (Liquid AI)","board":"api","source_url":"https://www.liquid.ai","axes":{"intelligence":69.70041168706942,"calibration":91.25367387086692,"speed":88.94430045323229,"cost":62.300793859384434},"capability":{"score":80.47704277896817,"rank":34,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":9},"composite_score":76.04601536374408,"rank":18,"open_board_rank":null,"api_board_rank":7,"measured_in":"v1.6.21","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-15","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01805677575757576,"basis":"Liquid AI public price list, USD 0.042/M input, output free (fetched 16 Oct 2026): the models table of Liquid's public pricing page lists d1-ultra at USD 0.042 per 1M input tokens and USD 0.00 per 1M output tokens (saved copy sha256 89fb8ef2...); the same rate Liquid's API billed on the measurement's smoke call (48 input tokens -> USD 0.000002016). Counted as the manufacturer's standard launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops and SubQ Snap), through a scoped re-freeze of the frozen pricing module that adds this one key and nothing else, with a zero-difference proof for every other registry entry. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (638,436 input tokens over the 1,485 answered items, 0 output tokens), divided by the 1,485 items of the common cost basis the scorer uses; the 15 items refused with HTTP 422 count as zero tokens and as wrong answers. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.018057 per 1,000 decisions (0.280x the official cost cap USD 0.064595), against USD 0.017876 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.2306590935040731,"p50_s_adjusted":0.2306590935040731,"adjustment":"none (API)","p50_s_measured":0.2306590935040731,"latency_basis":"api-end-to-end","p95_s_raw":0.5528398923954227,"p95_s_adjusted":0.5528398923954227,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":76.05614823638027,"capability":80.47704277896817,"axes":{"calibration":91.25367387086692,"cost":62.300793859384434,"intelligence":69.70041168706942,"speed":88.99977436791615},"rank":4,"open_board_rank":null,"api_board_rank":3}},{"key":"litjev","name":"LitJev (Qwen3.8-27B)","board":"open","source_url":"https://github.com/zhengxuyu/litjev","axes":{"intelligence":43.28549422942267,"calibration":89.9210925231847,"speed":68.06914373137408,"cost":55.72786321561281},"capability":{"score":66.60329337630368,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 5.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":44.83814187422011,"rank":151,"open_board_rank":119,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029904592776247543,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 944,931 input + 0 output tokens over 1,624 decisions = 581.854064039 tokens/decision. GPU-time cost USD 0.0299045927762475/1,000; old list-price cost USD 0.244378706896552/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0299045927762475/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":1.4977164075244218,"p50_s_adjusted":3.1454328150488435,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":1.4977164075244218,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.4045634037116543,"p95_s_adjusted":4.959126807423309,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":11.458886170296882,"capability":66.60329337630368,"axes":{"calibration":89.9210925231847,"cost":28.35809912639418,"intelligence":43.28549422942267,"speed":68.17671225113584},"rank":165,"open_board_rank":126,"api_board_rank":null}},{"key":"localjev-qwen3.5-4b","name":"local-jev Qwen3.5-4B","board":"open","source_url":"https://github.com/amithgc/local-jev","axes":{"intelligence":24.145363218466134,"calibration":85.84286362718517,"speed":85.6437855976805,"cost":77.53686751699048},"capability":{"score":54.99411342282565,"rank":201,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":167,"api_board_rank":null},"composite_score":12.014647872457989,"rank":243,"open_board_rank":196,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005607523247801365,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 925,135 input + 1,624 output tokens over 1,624 decisions = 570.664408867 tokens/decision. GPU-time cost USD 0.00560752324780137/1,000; old list-price cost USD 0.0227865763546798/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00560752324780137/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1373733568470925,"p50_s_adjusted":0.42474671369418504,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1373733568470925,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.24596776431426406,"p95_s_adjusted":0.6419355286285281,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":11.438161297684031,"capability":54.99411342282565,"axes":{"calibration":85.84286362718517,"cost":59.269627658320445,"intelligence":24.145363218466134,"speed":86.10068849774444},"rank":166,"open_board_rank":127,"api_board_rank":null}},{"key":"lumma-fev-27b","name":"FrontiersMind Lumma-fev-27B","board":"open","source_url":"https://huggingface.co/FrontiersMind/Lumma-fev-27B","axes":{"intelligence":55.45890006642501,"calibration":89.11821147102742,"speed":85.46346745302206,"cost":54.82196510185538},"capability":{"score":72.28855576872621,"rank":82,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":59,"api_board_rank":null},"composite_score":67.57419732927984,"rank":119,"open_board_rank":93,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03205785768536673,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 763,765 input + 162,504 output tokens over 1,485 decisions = 623.75016835 tokens/decision. GPU-time cost USD 0.0320578576853667/1,000; old list-price cost USD 0.544305252525253/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0320578576853667/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.14193624799997906,"p50_s_adjusted":0.43387249599995814,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.14193624799997906,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.25253758214998356,"p95_s_adjusted":0.6550751642999671,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"lychee-27b","name":"Lychee-27B","board":"open","source_url":"https://huggingface.co/LycheeAI/Lychee-27B","axes":{"intelligence":74.33387897478502,"calibration":85.36006033410777,"speed":83.81641932679139,"cost":57.33528640641721},"capability":{"score":79.8469696544464,"rank":37,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":28,"api_board_rank":null},"composite_score":73.34066782166333,"rank":55,"open_board_rank":40,"api_board_rank":null,"measured_in":"v1.7.20","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.20","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02643364905342198,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 763,765 input + 0 output tokens over 1,485 decisions = 514.31986532 tokens/decision. GPU-time cost USD 0.026433649053422/1,000; old list-price cost USD 0.216014343434343/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.026433649053422/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1770647059984185,"p50_s_adjusted":0.504129411996837,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1770647059984185,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3368945407000865,"p95_s_adjusted":0.823789081400173,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.20","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"malkuth-2b","name":"Malkuth-2B (newfull5, Kev post-train)","board":"open","source_url":"https://github.com/newfull5/malkuth","axes":{"intelligence":15.917395195670291,"calibration":79.34885520197874,"speed":91.96529998043799,"cost":83.82259492095744},"capability":{"score":47.633125198824516,"rank":226,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":190,"api_board_rank":null},"composite_score":4.126837854112484,"rank":260,"open_board_rank":211,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0034613605210043795,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 760,783 input + 169,020 output tokens over 1,624 decisions = 572.538793103 tokens/decision. GPU-time cost USD 0.00346136052100438/1,000; old list-price cost USD 0.0140538731527094/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00346136052100438/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04379177652299404,"p50_s_adjusted":0.23758355304598808,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04379177652299404,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05885162049380597,"p95_s_adjusted":0.26770324098761195,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.992169201455145,"capability":47.633125198824516,"axes":{"calibration":79.34885520197874,"cost":65.5661191189412,"intelligence":15.917395195670291,"speed":92.03462875001455},"rank":186,"open_board_rank":146,"api_board_rank":null}},{"key":"malkuth-4b","name":"Malkuth-4B (newfull5, Kev post-train)","board":"open","source_url":"https://github.com/newfull5/malkuth","axes":{"intelligence":33.840703805987175,"calibration":88.43741591513763,"speed":89.3735021970277,"cost":77.49131338115862},"capability":{"score":61.139059860562405,"rank":166,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":134,"api_board_rank":null},"composite_score":28.21050157509396,"rank":195,"open_board_rank":155,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005627163758940036,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 760,783 input + 169,222 output tokens over 1,624 decisions = 572.66317734 tokens/decision. GPU-time cost USD 0.00562716375894004/1,000; old list-price cost USD 0.0296839839901478/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00562716375894004/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07525279698893428,"p50_s_adjusted":0.3005055939778686,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07525279698893428,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11720612526114563,"p95_s_adjusted":0.3844122505222912,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":26.197476126697616,"capability":61.139059860562405,"axes":{"calibration":88.43741591513763,"cost":55.82433433950372,"intelligence":33.840703805987175,"speed":89.51571180654442},"rank":109,"open_board_rank":76,"api_board_rank":null}},{"key":"mammoth-1","name":"baselabs mammoth-1 (27B)","board":"open","source_url":"https://huggingface.co/baselabs/mammoth-1","axes":{"intelligence":76.20447479084162,"calibration":94.26152926942419,"speed":84.40895777972253,"cost":57.166365164587916},"capability":{"score":85.2330020301329,"rank":6,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":4,"api_board_rank":null},"composite_score":75.37036297447123,"rank":26,"open_board_rank":16,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.026778597517136134,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11840.1727884 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=base-arch). The row's own scored cost basis: 777,323 input + 1,485 output tokens over 1,485 decisions = 524.44983165 tokens/decision. GPU-time cost USD 0.0267785975171361/1,000; old list-price cost USD 0.103632717171717/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0267785975171361/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.18781221550000282,"p50_s_adjusted":0.5256244310000057,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.18781221550000282,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2696661899500214,"p95_s_adjusted":0.6893323799000428,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"manchego-v21-v15","name":"Manchego v2.1","board":"open","source_url":"https://github.com/nschlaepfer/manchego-serve","axes":{"intelligence":32.54303915979197,"calibration":84.0077997110931,"speed":90.57533514329342,"cost":78.87189352613305},"capability":{"score":58.27541943544253,"rank":176,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":144,"api_board_rank":null},"composite_score":25.53791020495781,"rank":204,"open_board_rank":163,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00506139481438631,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 836,500 input + 0 output tokens over 1,624 decisions = 515.086206897 tokens/decision. GPU-time cost USD 0.00506139481438631/1,000; old list-price cost USD 0.0154525862068966/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00506139481438631/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.055753487496986054,"p50_s_adjusted":0.26150697499397213,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.055753487496986054,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09247624424460808,"p95_s_adjusted":0.33495248848921616,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":24.487333472500133,"capability":58.27541943544253,"axes":{"calibration":84.0077997110931,"cost":64.32996474661243,"intelligence":32.54303915979197,"speed":90.7347510499745},"rank":114,"open_board_rank":81,"api_board_rank":null}},{"key":"mercury-decide","name":"Mercury Decide (Inception; System One decisions API, served free on OpenRouter as inception/mercury-decide:free)","board":"api","source_url":"https://openrouter.ai/inception/mercury-decide:free","axes":{"intelligence":65.90701345473542,"calibration":81.34341651168008,"speed":84.81995518783543,"cost":62.08313746631354},"capability":{"score":73.62521498320774,"rank":73,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":19},"composite_score":72.24968682544169,"rank":70,"open_board_rank":null,"api_board_rank":17,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.018360961408259987,"basis":"LABELLED ESTIMATE, not charged - and nothing WAS charged: usage.cost is 0 on every one of the 1,500 calls. Inception publishes NO paid tariff for Mercury Decide: OpenRouter lists it free-only (prompt 0, completion 0) and the non-free slug inception/mercury-decide has no endpoints at all (404 'No endpoints found', read 7 Oct 2026), so there is nothing bookable. Because the system is API-only with undisclosed weights, the open rows' size-class estimate table cannot apply either. The row is therefore priced from THE SAME OPERATOR'S cheapest published per-token tariff on the SAME platform, inception/mercury-2.5 at USD 0.04 per 1M input and USD 0.15 per 1M output, which IS a key in the frozen 25 Sep snapshot - x this row's own measured tokens. Output tokens are 3 per decision on average (4,434 over 1,477 answers) because the probability is read from the model rather than written out, so the output rate is nearly inert and the Cost axis is set by the input rate alone. PRICING REFERENCE NOTE: the rate inception/mercury-2.5 (0.04, 0.15) is already IN the dated snapshot; what is missing is only a base-model decision line pointing at it, so pricing_v15.reference() raises on every inception/* string. receipts/COST-ALTERNATIVES-R58.json has the two alternatives' arithmetic already done."},"latency":{"p50_s_raw":0.34386441484093666,"p50_s_adjusted":0.34386441484093666,"adjustment":"none (API)","p50_s_measured":0.34386441484093666,"latency_basis":"api-end-to-end","p95_s_raw":0.9585554625838959,"p95_s_adjusted":0.9585554625838959,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":72.36824049637728,"capability":73.62521498320774,"axes":{"calibration":81.34341651168008,"cost":62.08313746631354,"intelligence":65.90701345473542,"speed":85.47752366478576},"rank":15,"open_board_rank":null,"api_board_rank":11}},{"key":"messier-one-v0.2","name":"Messier One v0.2 (Qwen3.5-4B fine-tune, one prefill, self-hosted)","board":"open","source_url":"https://huggingface.co/agentmessier/messier-one","axes":{"intelligence":42.35139759222662,"calibration":73.90972729526267,"speed":92.56000105620268,"cost":79.42950661954919},"capability":{"score":58.130562443744644,"rank":177,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":145,"api_board_rank":null},"composite_score":47.407255888808336,"rank":145,"open_board_rank":113,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimated","usd_per_1000_decisions":0.004849345337009636,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 727,432 input + 1,477 output tokens over 1,477 decisions = 493.506431957 tokens/decision. GPU-time cost USD 0.00484934533700964/1,000; old list-price cost USD 0.0149251929587001/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00484934533700964/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.029460153993568383,"p50_s_adjusted":0.20892030798713676,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.029460153993568383,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05773615748912562,"p95_s_adjusted":0.26547231497825124,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":45.29059491870636,"capability":58.130562443744644,"axes":{"calibration":73.90972729526267,"cost":64.78240136302922,"intelligence":42.35139759222662,"speed":92.70909408602279},"rank":70,"open_board_rank":45,"api_board_rank":null}},{"key":"messier-one-v0.4","name":"Messier One v0.4 (Qwen3.5-4B fine-tune, self-hosted)","board":"open","source_url":"https://huggingface.co/agentmessier/messier-one","axes":{"intelligence":52.02542301918905,"calibration":72.86180997007926,"speed":89.42792459288492,"cost":75.66359272969224},"capability":{"score":62.443616494634156,"rank":152,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":120,"api_board_rank":null},"composite_score":69.75377301204921,"rank":101,"open_board_rank":79,"api_board_rank":null,"measured_in":"v1.6.22","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-14","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006474609433320888,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 978,475 input + 0 output tokens over 1,485 decisions = 658.905723906 tokens/decision. GPU-time cost USD 0.00647460943332089/1,000; old list-price cost USD 0.0197671717171717/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00647460943332089/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06901155365630984,"p50_s_adjusted":0.2880231073126197,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06901155365630984,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.1230387669522315,"p95_s_adjusted":0.396077533904463,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":66.16093508471093,"capability":62.443616494634156,"axes":{"calibration":72.86180997007926,"cost":61.121663950171616,"intelligence":52.02542301918905,"speed":89.67127141858641},"rank":45,"open_board_rank":24,"api_board_rank":null}},{"key":"metadecision-31b-rain","name":"MetaDecision-31B-Rain","board":"open","source_url":"https://huggingface.co/MeTask-Rain/MetaDecision-31B-Rain","axes":{"intelligence":85.04964777640174,"calibration":91.50530093288266,"speed":71.85456280157676,"cost":54.959198906753414},"capability":{"score":88.27747435464221,"rank":2,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":2,"api_board_rank":null},"composite_score":72.99472046195575,"rank":61,"open_board_rank":46,"api_board_rank":null,"measured_in":"v1.8.6","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.6","measured_on":"2026-10-27","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.031721961700337906,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 717,572 input + 72,041 output tokens over 1,489 decisions = 530.29751511 tokens/decision. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0317219617003379/1,000; old list-price estimate USD 0.0598223102753526/1,000 (google/gemma-4-31b-it via frozen pricing_v15 reference at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0317219617003379/1,000. Estimated, not a bill. USD 0.0317219617 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0329463697 over all 1,500 items)."},"latency":{"p50_s_raw":0.289625294999837,"p50_s_adjusted":0.7292505899996741,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.289625294999837,"latency_basis":"api-end-to-end","p95_s_raw":4.398391776250183,"p95_s_adjusted":8.946783552500367,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.6","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"metask-jev-4b","name":"metask-jev-4b","board":"open","source_url":"https://github.com/metask-ai/metask-jev","axes":{"intelligence":39.166454612935595,"calibration":84.41575352936586,"speed":90.45494495649334,"cost":76.02122363625783},"capability":{"score":61.79110407115073,"rank":157,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":125,"api_board_rank":null},"composite_score":39.85244736840569,"rank":173,"open_board_rank":136,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006299303678791922,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,041,090 input + 0 output tokens over 1,624 decisions = 641.065270936 tokens/decision. GPU-time cost USD 0.00629930367879192/1,000; old list-price cost USD 0.0256426108374384/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00629930367879192/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05803272299817763,"p50_s_adjusted":0.2660654459963553,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05803272299817763,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09423377123021055,"p95_s_adjusted":0.3384675424604211,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":37.34931073616377,"capability":61.79110407115073,"axes":{"calibration":84.41575352936586,"cost":57.7311327584882,"intelligence":39.166454612935595,"speed":90.78239008886905},"rank":87,"open_board_rank":57,"api_board_rank":null}},{"key":"metask-jev-rain-4b","name":"Metask rain 4B","board":"open","source_url":"https://huggingface.co/wayfind/metask-jev-4b-policy-mix","axes":{"intelligence":51.08240263971605,"calibration":80.89039994344657,"speed":78.38456523323399,"cost":69.50479145303603},"capability":{"score":65.9864012915813,"rank":128,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":101,"api_board_rank":null},"composite_score":67.70053221613982,"rank":117,"open_board_rank":92,"api_board_rank":null,"measured_in":"v1.6.2","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-09","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.010387402289587948,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,456,885 input + 106,568 output tokens over 1,479 decisions = 1057.10141988 tokens/decision. GPU-time cost USD 0.0103874022895879/1,000; old list-price cost USD 0.0403595334685598/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0103874022895879/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.08329759602202103,"p50_s_adjusted":0.3165951920440421,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08329759602202103,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.215916002256563,"p95_s_adjusted":4.581832004513126,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":62.69717160365172,"capability":65.9864012915813,"axes":{"calibration":80.89039994344657,"cost":51.82161587751452,"intelligence":51.08240263971605,"speed":79.59790373470649},"rank":52,"open_board_rank":31,"api_board_rank":null}},{"key":"metask_jev_rain_12b","name":"metask-jev-rain-12B","board":"open","source_url":"https://huggingface.co/cyankiwi/gemma-4-12B-it-qat-AWQ-INT4","axes":{"intelligence":66.74877414880126,"calibration":89.84415618532353,"speed":75.19061343032087,"cost":65.2582052479506},"capability":{"score":78.2964651670624,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":73.07910352308468,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.2","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-09","status":"anchor-runs-pending"},"ranked":false,"listing":"wrapper","not_ranked_because":"registry ranked=false or no composite","price":{"kind":"estimate","usd_per_1000_decisions":0.014389967897360545,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 859,455 input + 86,179 output tokens over 1,479 decisions = 639.373901285 tokens/decision. GPU-time cost USD 0.0143899678973605/1,000; old list-price cost USD 0.037795537525355/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0143899678973605/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10124877700582147,"p50_s_adjusted":0.35249755401164296,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10124877700582147,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":4.217917609983124,"p95_s_adjusted":8.585835219966249,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.66306996476943,"capability":78.2964651670624,"axes":{"calibration":89.84415618532353,"cost":52.6767842132323,"intelligence":66.74877414880126,"speed":75.9887043236542},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"metask_jev_rain_12b_trained","name":"metask-jev-rain-12B-trained","board":"open","source_url":"https://huggingface.co/MeTask-Rain/metask-jev-rain-12B-trained","axes":{"intelligence":66.48910229704197,"calibration":87.29383335481249,"speed":71.7899932169731,"cost":65.18480974745836},"capability":{"score":76.89146782592724,"rank":57,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":42,"api_board_rank":null},"composite_score":71.72811700838548,"rank":80,"open_board_rank":63,"api_board_rank":null,"measured_in":"v1.6.10","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.014471259844852572,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 870,611 input + 84,223 output tokens over 1,485 decisions = 642.985858586 tokens/decision. GPU-time cost USD 0.0144712598448526/1,000; old list-price cost USD 0.0378208754208754/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0144712598448526/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12866936152568087,"p50_s_adjusted":0.40733872305136176,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12866936152568087,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":8.05358562800684,"p95_s_adjusted":16.257171256013677,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":12,"n_over_limit_answered_ok":12,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.10451633937213,"capability":77.87178459131476,"axes":{"calibration":87.91805023499681,"cost":52.66805268932843,"intelligence":67.8255189476327,"speed":73.37579625429389},"rank":37,"open_board_rank":18,"api_board_rank":null}},{"key":"metask_jev_rain_31b_trained","name":"metask-jev-rain-31B-trained","board":"open","source_url":"https://huggingface.co/MeTask-Rain/metask-jev-rain-31B-trained","axes":{"intelligence":70.71905450578407,"calibration":87.09471194592824,"speed":82.33528699876703,"cost":56.13017070777345},"capability":{"score":78.90688322585615,"rank":43,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":33,"api_board_rank":null},"composite_score":71.96392483986129,"rank":77,"open_board_rank":60,"api_board_rank":null,"measured_in":"v1.6.10","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.028995302439358734,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 718,318 input + 1,485 output tokens over 1,485 decisions = 484.715824916 tokens/decision. GPU-time cost USD 0.0289953024393587/1,000; old list-price cost USD 0.0438744242424242/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0289953024393587/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.2330208719940856,"p50_s_adjusted":0.6160417439881712,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2330208719940856,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.39905765381990926,"p95_s_adjusted":0.9481153076398185,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":69.74074308829535,"capability":78.90688322585615,"axes":{"calibration":87.09471194592824,"cost":50.73365709012802,"intelligence":70.71905450578407,"speed":83.17884674032656},"rank":28,"open_board_rank":12,"api_board_rank":null}},{"key":"metask_jev_rain_31b_trained_v3","name":"metask-jev-rain-31B-trained service v3","board":"open","source_url":"https://huggingface.co/MeTask-Rain/metask-jev-rain-31B-trained","axes":{"intelligence":72.59243765084337,"calibration":86.60202339920977,"speed":73.27177592289428,"cost":55.40100072278519},"capability":{"score":79.59723052502656,"rank":39,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":29,"api_board_rank":null},"composite_score":70.15002550079586,"rank":96,"open_board_rank":74,"api_board_rank":null,"measured_in":"v1.6.17","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03066431794216985,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 743,854 input + 17,382 output tokens over 1,485 decisions = 512.616835017 tokens/decision. GPU-time cost USD 0.0306643179421698/1,000; old list-price cost USD 0.0490617777777778/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0306643179421698/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1847877809777856,"p50_s_adjusted":0.5195755619555712,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1847877809777856,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":4.455474559939465,"p95_s_adjusted":9.06094911987893,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.6558274201148,"capability":79.59723052502656,"axes":{"calibration":86.60202339920977,"cost":49.27770156907625,"intelligence":72.59243765084337,"speed":84.66561996644248},"rank":41,"open_board_rank":21,"api_board_rank":null}},{"key":"mghafiri-qwen3.5-0.8b-decision-model","name":"Qwen3.5-0.8B Decision Model (Mourad Ghafiri)","board":"open","source_url":"https://huggingface.co/mghafiri/qwen3.5-0.8B-decision-model","axes":{"intelligence":7.297270616183145,"calibration":77.95657740696296,"speed":55.90747410573378,"cost":86.07480552202115},"capability":{"score":42.62692401157305,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 18.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.47499805397335026,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":false,"listing":"remeasure-pending","not_ranked_because":"Re-measure pending under the 10 s rule: this row's per-decision times were recorded on a loaded CPU host, so they cannot be scored under the time limit. Shown with its v1.6 values until it is measured again.","price":{"kind":"estimate","usd_per_1000_decisions":0.0029118778259148497,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 782,199 input + 0 output tokens over 1,624 decisions = 481.649630542 tokens/decision. GPU-time cost USD 0.00291187782591485/1,000; old list-price cost USD 0.00481649630541872/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00291187782591485/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":5.551045611966401,"p50_s_adjusted":11.252091223932803,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":5.551045611966401,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":11.327218242688105,"p95_s_adjusted":22.804436485376208,"n":124},"method":"v1.6","rescored_in":null,"method_status":"v1.6-pending","decision_limit":null,"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.47247471382875306,"capability":42.62692401157305,"axes":{"calibration":77.95657740696296,"cost":79.5180630553356,"intelligence":7.297270616183145,"speed":55.90747410573378},"rank":207,"open_board_rank":164,"api_board_rank":null}},{"key":"microsoft-decision-1","name":"Microsoft-Decision-1 (Azure Foundry)","board":"api","source_url":"https://commandline.microsoft.com/microsoft-decision-1-model-foundry/","axes":{"intelligence":57.16958510591244,"calibration":84.39873071167106,"speed":81.38546065308665,"cost":61.452075716621465},"capability":{"score":70.78415790879176,"rank":96,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":23},"composite_score":69.08416559129289,"rank":103,"open_board_rank":null,"api_board_rank":23,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.0192721784989858,"basis":"Advertised Azure Foundry tariff .042/M input, output free; genuine native usage, billed invoice unknown"},"latency":{"p50_s_raw":0.4787515588104725,"p50_s_adjusted":0.4787515588104725,"adjustment":"none (API)","p50_s_measured":0.4787515588104725,"latency_basis":"api-end-to-end","p95_s_raw":1.5182515516877173,"p95_s_adjusted":1.5182515516877173,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":69.11289175277149,"capability":70.78415790879176,"axes":{"calibration":84.39873071167106,"cost":61.452075716621465,"intelligence":57.16958510591244,"speed":81.54517543213291},"rank":29,"open_board_rank":null,"api_board_rank":17}},{"key":"microsoft-decision-1-r2","name":"Microsoft-Decision-1 (updated weights, Oct 2026)","board":"api","source_url":"https://commandline.microsoft.com/microsoft-decision-1-model-foundry/","axes":{"intelligence":64.22537211706586,"calibration":90.64593876574243,"speed":89.34376153841384,"cost":66.40926993775184},"capability":{"score":77.43565544140415,"rank":52,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":14},"composite_score":75.68165551195851,"rank":23,"open_board_rank":null,"api_board_rank":9,"measured_in":"v1.7.15","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.15","measured_on":"2026-10-28","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.013173191403626596,"basis":"Advertised Azure Foundry tariff USD 0.042 per million input tokens, output free (same basis as the first Microsoft-Decision-1 row); provider-reported usage. COST FIGURES: USD 0.013173 per 1,000 decisions, 0.20x the official cost cap USD 0.064595. Cost on the 1,489 items every row on this draw answered (same-items rule); all-items cost USD 0.013138 per 1,000 decisions. Over all n = 1,493 decisions: 467,021 input and 1,489 output tokens, the 4 refused requests in the denominator. Billed invoice unknown."},"latency":{"p50_s_raw":0.31938244198681787,"p50_s_adjusted":0.31938244198681787,"adjustment":"none (API)","p50_s_measured":0.31938244198681787,"latency_basis":"api-end-to-end","p95_s_raw":0.3641772473813034,"p95_s_adjusted":0.3641772473813034,"n":1493},"method":"v1.7","rescored_in":null,"new_in":"v1.7.15","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1493,"treatment":"share"},"previous":null},{"key":"minimodels-decision-1.0-candy-cane","name":"Minimodels Decision 1.0 (candy-cane)","board":"open","source_url":"https://huggingface.co/minimodels-ai/decision-1.0","axes":{"intelligence":74.2250736791645,"calibration":90.7247525637117,"speed":72.48013389204425,"cost":57.60805349713269},"capability":{"score":82.4749131214381,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":71.87722052632512,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.11","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.11","measured_on":"2026-11-15","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.02588599622055449,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.0258859962205545/1,000; old list-price estimate USD 0.0409948891873741/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0258859962205545/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.1296297365000214,"p50_s_adjusted":0.40925947300004284,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1296297365000214,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":6.826728554449987,"p95_s_adjusted":13.803457108899975,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.11","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"mirror","name":"Mirror","board":"open","source_url":"https://github.com/Bluusun/Decision-API","axes":{"intelligence":0,"calibration":51.61277239297377,"speed":74.03935684279031,"cost":89.27300805786774},"capability":{"score":25.806386196486883,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 2.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0,"rank":313,"open_board_rank":260,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0022780665024630543,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.5608240149449557,"p50_s_adjusted":1.2716480298899113,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.5608240149449557,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.476198575645685,"p95_s_adjusted":3.10239715129137,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0,"capability":25.806386196486883,"axes":{"calibration":51.61277239297377,"cost":89.27300805786774,"intelligence":0,"speed":73.47211485993975},"rank":240,"open_board_rank":196,"api_board_rank":null}},{"key":"modernbert-jev-v2","name":"modernbert-jev v2","board":"open","source_url":"https://huggingface.co/ksonny4/modernbert-jev-v2","axes":{"intelligence":2.6107469083924393,"calibration":66.68251917899094,"speed":57.79924670989358,"cost":70.36992715953406},"capability":{"score":34.64663304369169,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 8.7× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.025388991293562355,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.8","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":false,"listing":"remeasure-pending","not_ranked_because":"Re-measure pending under the 10 s rule: this row's per-decision times were recorded on a loaded CPU host, so they cannot be scored under the time limit. Shown with its v1.6 values until it is measured again.","price":{"kind":"estimate","usd_per_1000_decisions":0.009720063333333334,"basis":"ESTIMATE (*est.), size-class reference; self-hosted open weights, no public tariff. The same choice the published quyet-1-0-small-en row carries for the identical ModernBERT-base backbone: DeepInfra base-size encoders at USD 0.005 per 1M input and 0 output (deepinfra:thenlper/gte-base, BAAI/bge-base-en-v1.5, intfloat/e5-base-v2 and sentence-transformers/all-mpnet-base-v2 are all 5e-7 cents per input token in the frozen 25 Sep 2026 DeepInfra snapshot). Output tokens are a literal 0 reported by the author's server (NLI entailment read-out, nothing is generated). This row's own usage on draw v1.6-regular-20261010-a2: 2,916,019 input and 0 output tokens over 1,500 decisions. Estimated, not charged."},"latency":{"p50_s_raw":2.5996471606194973,"p50_s_adjusted":5.349294321238995,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":2.5996471606194973,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":15.439894960820675,"p95_s_adjusted":31.02978992164135,"n":124},"method":"v1.6","rescored_in":null,"method_status":"v1.6-pending","decision_limit":null,"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.025388991293562355,"capability":34.64663304369169,"axes":{"calibration":66.68251917899094,"cost":70.36992715953406,"intelligence":2.6107469083924393,"speed":57.79924670989358},"rank":229,"open_board_rank":185,"api_board_rank":null}},{"key":"mxbai-rerank-base-v2","name":"Mixedbread mxbai-rerank-base-v2","board":"open","source_url":"https://huggingface.co/mixedbread-ai/mxbai-rerank-base-v2","axes":{"intelligence":1.140707327109041,"calibration":86.60772459153094,"speed":92.05600895369167,"cost":60.33589314238483},"capability":{"score":43.87421595931999,"rank":237,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":200,"api_board_rank":null},"composite_score":0.0022737744722840765,"rank":307,"open_board_rank":254,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020996016009852216,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: size-class proxy (deepinfra:Qwen/Qwen3-Reranker-0.6B); no exact base-model floor applies"},"latency":{"p50_s_raw":0.03210197016596794,"p50_s_adjusted":0.21420394033193588,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03210197016596794,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0703923825640231,"p95_s_adjusted":0.2907847651280462,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.002273929358844764,"capability":43.87421595931999,"axes":{"calibration":86.60772459153094,"cost":60.33589314238483,"intelligence":1.140707327109041,"speed":92.58758170313658},"rank":234,"open_board_rank":190,"api_board_rank":null}},{"key":"nagi-enormous","name":"Nagi-ENORMOUS 27B (nagisanzenin, Qwen3.8-27B LoRA)","board":"open","source_url":"https://huggingface.co/nagisanzeninz/Nagi-ENORMOUS","axes":{"intelligence":62.38818589248167,"calibration":85.14254261258804,"speed":85.82001330475426,"cost":57.35797715995501},"capability":{"score":73.76536425253485,"rank":72,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":53,"api_board_rank":null},"composite_score":70.34779585768042,"rank":93,"open_board_rank":72,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02638765281165652,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 762,436 input + 0 output tokens over 1,485 decisions = 513.424915825 tokens/decision. GPU-time cost USD 0.0263876528116565/1,000; old list-price cost USD 0.215638464646465/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0263876528116565/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15377032081596553,"p50_s_adjusted":0.4575406416319311,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15377032081596553,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2111139261862263,"p95_s_adjusted":0.5722278523724526,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":19.784711166742767,"capability":73.76536425253485,"axes":{"calibration":85.14254261258804,"cost":29.98821307073638,"intelligence":62.38818589248167,"speed":86.16105911802293},"rank":132,"open_board_rank":97,"api_board_rank":null}},{"key":"nagi-huge","name":"Nagi-HUGE 12B (nagisanzenin, Gemma-4-12B LoRA)","board":"open","source_url":"https://huggingface.co/nagisanzeninz/Nagi-HUGE","axes":{"intelligence":54.80664491424396,"calibration":73.2732669877037,"speed":88.38293814019335,"cost":null},"capability":{"score":64.03995595097383,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: google/gemma-4-12B-it is in the frozen module's UNLISTED_BASE_MODELS (no eligible base-model price); one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 3948e3f1..., and the later re-freezes, which only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base google/gemma-4-12B-it (the module lists it in UNLISTED_BASE_MODELS, so reference() returns no price); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens as reported by the evaluator's wrapper (attention-mask ones over every forward pass: input 764,280, output_tokens 0)."},"latency":{"p50_s_raw":0.10029390454292297,"p50_s_adjusted":0.35058780908584597,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10029390454292297,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.13195664717350153,"p95_s_adjusted":0.4139132943470031,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":64.03995595097383,"axes":{"calibration":73.2732669877037,"cost":null,"intelligence":54.80664491424396,"speed":88.62940634950104},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"nemotron-deciffusion-8b-v2","name":"Nemotron Deciffusion 8B v2 (pst2154)","board":"open","source_url":"https://huggingface.co/nemotron-deciffusion/Nemotron-Deciffusion-8B-v2","axes":{"intelligence":49.60038985253669,"calibration":89.03920345557668,"speed":87.35910345792061,"cost":null},"capability":{"score":69.31979665405669,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.22","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-15","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base nvidia/Nemotron-Labs-Diffusion-8B and raises rather than estimating; a future published pricing update may add an entry","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module RAISES on the declared base nvidia/Nemotron-Labs-Diffusion-8B ('base model lacks an explicit market-listing decision') and the dated snapshot carries no listing of that family, so this row carries no pricing_base and is listed, measured and shown WITHOUT a cost number and WITHOUT a rank - the pre-declared default (Sage 1.4, Celeris Option A). No rate is invented. The predecessor nemotron-diffusion-8b row's usd_per_1000_override is a v1.5.4 historical carry on a different draw and is deliberately NOT reused. Pricing remains unresolved; the server reports both token fields, so a later pricing decision needs no re-measurement."},"latency":{"p50_s_raw":0.10529770236462355,"p50_s_adjusted":0.3605954047292471,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10529770236462355,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.17970617301762104,"p95_s_adjusted":0.5094123460352421,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":69.31979665405669,"axes":{"calibration":89.03920345557668,"cost":null,"intelligence":49.60038985253669,"speed":87.70433528726323},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"nemotron-diffusion-8b","name":"Nemotron Diffusion 8B (pst2154, optimized vLLM)","board":"open","source_url":"https://github.com/pst2154/Nemotron_Jev","axes":{"intelligence":20.38848344571935,"calibration":74.53740737677316,"speed":93.55093866077343,"cost":55.49170925809556},"capability":{"score":47.46294541124625,"rank":228,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":192,"api_board_rank":null},"composite_score":7.294934989446692,"rank":255,"open_board_rank":207,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03045157019704434,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.024202676489949226,"p50_s_adjusted":0.19840535297989845,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.024202676489949226,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.036255823587998746,"p95_s_adjusted":0.22251164717599747,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":7.29748476516348,"capability":47.46294541124625,"axes":{"calibration":74.53740737677316,"cost":55.49170925809556,"intelligence":20.38848344571935,"speed":93.8305724848098},"rank":179,"open_board_rank":140,"api_board_rank":null}},{"key":"nev-2b","name":"Nev-2B (EryriLabs)","board":"open","source_url":"https://huggingface.co/EryriLabs/Nev-2B-GGUF","axes":{"intelligence":20.23194846537782,"calibration":79.60239191738147,"speed":54.82852093556197,"cost":86.62542366165748},"capability":{"score":49.91717019137964,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 20.8× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":7.136510525770009,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.22","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-15","status":"anchor-runs-pending"},"ranked":false,"listing":"remeasure-pending","not_ranked_because":"Re-measure pending under the 10 s rule: this row's per-decision times were recorded on a loaded CPU host, so they cannot be scored under the time limit. Shown with its v1.6 values until it is measured again.","price":{"kind":"estimate","usd_per_1000_decisions":0.002791381599059377,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 685,652 input + 0 output tokens over 1,485 decisions = 461.718518519 tokens/decision. GPU-time cost USD 0.00279138159905938/1,000; old list-price cost USD 0.00923437037037037/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00279138159905938/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":6.343097407487221,"p50_s_adjusted":12.836194814974442,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":6.343097407487221,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":12.738908278940652,"p95_s_adjusted":25.627816557881303,"n":124},"method":"v1.6","rescored_in":null,"method_status":"v1.6-pending","decision_limit":null,"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":6.944821492706128,"capability":49.91717019137964,"axes":{"calibration":79.60239191738147,"cost":71.03778132505249,"intelligence":20.23194846537782,"speed":54.82852093556197},"rank":181,"open_board_rank":141,"api_board_rank":null}},{"key":"nimble-9b","name":"Bespoke Nimble 9B (Bespoke Labs)","board":"open","source_url":"https://github.com/bespokelabsai/nimble","axes":{"intelligence":43.14237010986392,"calibration":69.88762704651819,"speed":85.53966579201045,"cost":56.42334655643978},"capability":{"score":56.514998578191054,"rank":191,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":157,"api_board_rank":null},"composite_score":44.51360799306354,"rank":152,"open_board_rank":120,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02835013392240557,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,079,172 input + 3,248 output tokens over 1,624 decisions = 1282.27832512 tokens/decision. GPU-time cost USD 0.0283501339224056/1,000; old list-price cost USD 0.128327832512315/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0283501339224056/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15715505019761622,"p50_s_adjusted":0.46431010039523246,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15715505019761622,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2257428381592035,"p95_s_adjusted":0.601485676318407,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.082616704957896,"capability":56.514998578191054,"axes":{"calibration":69.88762704651819,"cost":36.75037423027614,"intelligence":43.14237010986392,"speed":86.11327010751721},"rank":124,"open_board_rank":90,"api_board_rank":null}},{"key":"nimble-v4","name":"Bespoke Nimble-v4 (Bespoke Labs)","board":"api","source_url":"https://bespokelabs.ai","axes":{"intelligence":41.405096245385266,"calibration":86.63426567315432,"speed":88.52559981079375,"cost":60.74282184772803},"capability":{"score":64.01968095926979,"rank":138,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":29},"composite_score":43.22874272615112,"rank":162,"open_board_rank":null,"api_board_rank":34,"measured_in":"v1.6.18","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-14","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020350383838383838,"basis":"Bespoke Labs public list price, USD 0.04/M input, output free (fetched 16 Oct 2026): Bespoke's Nimble pricing page (docs.bespokelabs.ai/nimble/pricing, saved copy sha256 03b06261...) says \"Nimble charges for input tokens only. Output tokens are free.\" and lists nimble-v4 at USD 0.04 per 1M input tokens. Counted as the manufacturer's standard, non-promotional launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap and d1-ultra), through a scoped re-freeze of the frozen pricing module that adds exactly three keys (sage-1.4.0, nimble-v4, nimble-v4-max) to its launch-list-price set and nothing else, with a zero-difference proof for every other registry entry and the independent release review; it is the separate pricing change made the cost axis wait for. Free credit for new accounts is a trial credit and is not used: the row is priced at the published tariff, never as cost 0. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (1,935,351 input tokens over all 1,500 answered items; the 3,000 reported output tokens are free under the tariff), divided by the 1,485 items of the common cost basis the scorer uses. USD 0.020350 per 1,000 decisions on the common basis (USD 0.051609 over all 1,500 items), 0.315x the official cost cap USD 0.064595. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.020350 per 1,000 decisions (0.315x the official cost cap USD 0.064595), against USD 0.051609 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.2323382355098147,"p50_s_adjusted":0.2323382355098147,"adjustment":"none (API)","p50_s_measured":0.2323382355098147,"latency_basis":"api-end-to-end","p95_s_raw":0.6043928697079537,"p95_s_adjusted":0.6043928697079537,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":43.21768647452467,"capability":64.01968095926979,"axes":{"calibration":86.63426567315432,"cost":60.74282184772803,"intelligence":41.405096245385266,"speed":88.39856712562789},"rank":74,"open_board_rank":null,"api_board_rank":26}},{"key":"nimble-v4-max","name":"Bespoke Nimble-v4-max (Bespoke Labs)","board":"api","source_url":"https://bespokelabs.ai","axes":{"intelligence":68.18491883484498,"calibration":91.24998737212707,"speed":87.92708417376494,"cost":60.74282184772803},"capability":{"score":79.71745310348602,"rank":38,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":10},"composite_score":74.82115516763538,"rank":31,"open_board_rank":null,"api_board_rank":11,"measured_in":"v1.6.18","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-14","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020350383838383838,"basis":"Bespoke Labs public list price, USD 0.04/M input, output free: Bespoke's Nimble pricing page (docs.bespokelabs.ai/nimble/pricing) says \"Nimble charges for input tokens only. Output tokens are free.\" and, in the copy fetched 16 Oct 2026 at about 21:40Z (saved copy sha256 4d46184d...), lists nimble-v4-max at USD 0.04 per 1M input tokens. The copy of the same page fetched 16 Oct 2026 at 15:58Z (saved copy sha256 03b06261...), from which v1.6.23 first priced this row, listed USD 0.09; v1.6.26 re-prices the row to the new public tariff through a scoped re-freeze of the frozen pricing module that changes the documented tariff of this one key and nothing else (no key joins or leaves its launch-list-price set; zero difference for every other registry entry). This is the current public list price on the vendor's pricing page (changed 16 Oct 2026; earlier copy listed USD 0.09), counted under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap and d1-ultra). Free credit for new accounts is a trial credit and is not used: the row is priced at the published tariff, never as cost 0. No new measurement: ESTIMATE in the scorer's sense, the tariff x this row's OWN provider-reported tokens from its v1.6.18 measurement (1,935,351 input tokens over all 1,500 answered items; the 3,000 reported output tokens are free under the tariff), divided by the 1,485 items of the common cost basis the scorer uses. USD 0.020350 per 1,000 decisions on the common basis (USD 0.051609 over all 1,500 items), 0.315x the official cost cap USD 0.064595 (at the earlier USD 0.09 tariff it was over the cap). Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.020350 per 1,000 decisions (0.315x the official cost cap USD 0.064595), against USD 0.051609 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.2574840509914793,"p50_s_adjusted":0.2574840509914793,"adjustment":"none (API)","p50_s_measured":0.2574840509914793,"latency_basis":"api-end-to-end","p95_s_raw":0.6259523141590762,"p95_s_adjusted":0.6259523141590762,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":75.4395528754179,"capability":79.71745310348602,"axes":{"calibration":91.24998737212707,"cost":60.74282184772803,"intelligence":68.18491883484498,"speed":91.45091520080334},"rank":6,"open_board_rank":null,"api_board_rank":5}},{"key":"nimble-v4-max-open","name":"Nimble-v4-max (open weights, self-hosted)","board":"open","source_url":"https://huggingface.co/bespokelabs/Nimble-v4-max","axes":{"intelligence":70.95116002556868,"calibration":91.36294095928801,"speed":84.78525949557437,"cost":57.47690696484853},"capability":{"score":81.15705049242834,"rank":29,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":21,"api_board_rank":null},"composite_score":73.75670981544037,"rank":48,"open_board_rank":35,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.026147877068277196,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 755,508 input + 0 output tokens over 1,485 decisions = 508.75959596 tokens/decision. GPU-time cost USD 0.0261478770682772/1,000; old list-price cost USD 0.21367903030303/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0261478770682772/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15395507599987468,"p50_s_adjusted":0.4579101519997494,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15395507599987468,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.287797079800157,"p95_s_adjusted":0.725594159600314,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"ninfer-qwen3.8-27b","name":"NInfer Qwen3.8-27B NVFP4","board":"open","source_url":"https://github.com/igorls/ninfer","axes":{"intelligence":51.800126951055375,"calibration":84.18557302696732,"speed":90.16889017946136,"cost":56.4886519334233},"capability":{"score":67.99284998901135,"rank":121,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":94,"api_board_rank":null},"composite_score":66.69283025993185,"rank":121,"open_board_rank":95,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02820838801735135,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 891,334 input + 0 output tokens over 1,624 decisions = 548.850985222 tokens/decision. GPU-time cost USD 0.0282083880173514/1,000; old list-price cost USD 0.230517413793103/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0282083880173514/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05742834598640911,"p50_s_adjusted":0.26485669197281825,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05742834598640911,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10658084631140809,"p95_s_adjusted":0.36316169262281617,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.71600496541158,"capability":67.99284998901135,"axes":{"calibration":84.18557302696732,"cost":29.118887844204664,"intelligence":51.800126951055375,"speed":90.47734239273063},"rank":143,"open_board_rank":107,"api_board_rank":null}},{"key":"ninfer-qwen3.8-flash-next","name":"NInfer Qwen3.8-Flash-Next mixed","board":"open","source_url":"https://github.com/igorls/ninfer","axes":{"intelligence":48.980301026876646,"calibration":90.37232453477458,"speed":89.03327048907914,"cost":71.76650263280871},"capability":{"score":69.67631278082561,"rank":108,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":82,"api_board_rank":null},"composite_score":67.76131625061458,"rank":116,"open_board_rank":91,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00873205734036537,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 891,334 input + 0 output tokens over 1,624 decisions = 548.850985222 tokens/decision. GPU-time cost USD 0.00873205734036537/1,000; old list-price cost USD 0.0823276477832512/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00873205734036537/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.08449016883969307,"p50_s_adjusted":0.31898033767938616,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08449016883969307,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12082991795614362,"p95_s_adjusted":0.39165983591228726,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":41.96785076957165,"capability":69.67631278082561,"axes":{"calibration":90.37232453477458,"cost":42.53362878447125,"intelligence":48.980301026876646,"speed":89.33173525631216},"rank":78,"open_board_rank":51,"api_board_rank":null}},{"key":"noma","name":"Noma (Blackdrome AI Labs, Qwen3.5-4B-Base derivative)","board":"open","source_url":"https://huggingface.co/BlackdromeAILabs/noma","axes":{"intelligence":25.051442411072966,"calibration":83.63708212673413,"speed":93.07972984765502,"cost":76.49547138218298},"capability":{"score":54.34426226890355,"rank":205,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":171,"api_board_rank":null},"composite_score":13.266137669977962,"rank":240,"open_board_rank":193,"api_board_rank":null,"measured_in":"v1.6.24","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006074132863021752,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 917,953 input + 0 output tokens over 1,485 decisions = 618.15016835 tokens/decision. GPU-time cost USD 0.00607413286302175/1,000; old list-price cost USD 0.0185445050505051/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00607413286302175/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02593987550062593,"p50_s_adjusted":0.20187975100125186,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02593987550062593,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.046872090056422164,"p95_s_adjusted":0.24374418011284432,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":12.754183690755768,"capability":54.34426226890355,"axes":{"calibration":83.63708212673413,"cost":61.95354260266235,"intelligence":25.051442411072966,"speed":93.34244428975083},"rank":163,"open_board_rank":124,"api_board_rank":null}},{"key":"nvidia/nemotron-3.5-super-vl","name":"NVIDIA Nemotron 3.5 Super VL","board":"api","source_url":"https://openrouter.ai/nvidia/nemotron-3.5-super-vl","axes":{"intelligence":57.489667803442686,"calibration":56.90864428042607,"speed":81.97747198900981,"cost":47.22136754775445},"capability":{"score":57.19915604193437,"rank":181,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":33},"composite_score":52.20484579181971,"rank":141,"open_board_rank":null,"api_board_rank":31,"measured_in":"v1.8.7","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.7","measured_on":"2026-11-03","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.05744969778374748,"basis":"Published OpenRouter tariff for the Crusoe bf16 endpoint of nvidia/nemotron-3.5-super-vl: USD 0.10 per million input tokens and USD 0.40 per million output tokens (fetched 2026-11-03T05:45:33Z; frozen pricing module sha256 2f52370fe1dba73d75a770a386df350cc76a0898304aeba9e29c91f6e4d3738b), counted from day 1 as the operator launch list price under the standing launch-list-price interpretation (I-1; the authority is recorded in the release proof). Hosted list price, so no GPU-time estimate applies. Cost = tariff x the engine-reported usage on the frozen 1,489-item common set of this draw: 607,138 input and 62,072 output tokens = 449.435863 tokens/decision. COST FIGURES: USD 0.057450 per 1,000 decisions. Cost on the 1,489 items every row on this draw answered (same-items rule); all-items cost USD 0.084429 per 1,000 decisions over all 1,500 decisions (1,015,566 input and 62,716 output tokens, reasoning disabled so no reasoning tokens). Tariff-based figure from engine-reported usage, not a bill."},"latency":{"p50_s_raw":0.6272735850070603,"p50_s_adjusted":0.6272735850070603,"adjustment":"none (API)","p50_s_measured":0.6272735850070603,"latency_basis":"api-end-to-end","p95_s_raw":1.0111039375420658,"p95_s_adjusted":1.0111039375420658,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.7","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"oenogarum","name":"oenogarum (plantwaterer; commercially restricted)","board":"open","source_url":"https://huggingface.co/plantwaterer/oenogarum","axes":{"intelligence":71.80689210610204,"calibration":91.25116778108405,"speed":84.93738695441665,"cost":55.89119947180174},"capability":{"score":81.52902994359305,"rank":26,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":19,"api_board_rank":null},"composite_score":73.32683753712675,"rank":56,"open_board_rank":41,"api_board_rank":null,"measured_in":"v1.8.2","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.2","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029532033400877865,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 735,102 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0295320334008779/1,000; old list-price estimate USD 0.0444319543317663/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0295320334008779/1,000. Estimated, not a bill. USD 0.0295320334 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0295061687 over all 1,500 items)."},"latency":{"p50_s_raw":0.14251593699987097,"p50_s_adjusted":0.43503187399974197,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.14251593699987097,"latency_basis":"api-end-to-end","p95_s_raw":0.29373149209997107,"p95_s_adjusted":0.7374629841999422,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.2","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"open-alternative-jev","name":"open-alternative-jev (Qwen3.5-4B, IkerMoel)","board":"open","source_url":"https://github.com/ikermoel/open-alternative-jev","axes":{"intelligence":22.34610151162883,"calibration":75.94384178347806,"speed":90.86432704430521,"cost":63.27859776193043},"capability":{"score":49.14497164755345,"rank":222,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":186,"api_board_rank":null},"composite_score":9.429815011124056,"rank":247,"open_board_rank":200,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0167512315270936,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (proxy tokens, I-2): exact base-model market reference; reference deepinfra:Qwen/Qwen3.5-4B frozen at the 25 Sep cut-off; the snapshot records deprecation on 11 Jun 2026 and replacement by Qwen/Qwen3.5-9B; the frozen snapshot rates remain in use (I-3)"},"latency":{"p50_s_raw":0.050254593981662765,"p50_s_adjusted":0.25050918796332555,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.050254593981662765,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.08857376150263008,"p95_s_adjusted":0.3271475230052602,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":9.435692571873393,"capability":49.14497164755345,"axes":{"calibration":75.94384178347806,"cost":63.27859776193043,"intelligence":22.34610151162883,"speed":91.30216983336405},"rank":173,"open_board_rank":134,"api_board_rank":null}},{"key":"open-jev-deberta-v3-large","name":"open-jev-deberta-v3-large (local CPU)","board":"open","source_url":"https://github.com/kotoba-lang/typed-decisions","axes":{"intelligence":2.8293414417857914,"calibration":67.96597988813339,"speed":67.31985953521365,"cost":77.5922354035203},"capability":{"score":35.39766066495959,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 5.7× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.03235273404518239,"rank":302,"open_board_rank":249,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005583743842364532,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: v1.5 measured-input proxy; no exact base-model floor applies"},"latency":{"p50_s_raw":1.6850177072919905,"p50_s_adjusted":3.520035414583981,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":1.6850177072919905,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.5579160863999277,"p95_s_adjusted":5.265832172799856,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.03235814777127837,"capability":35.39766066495959,"axes":{"calibration":67.96597988813339,"cost":77.5922354035203,"intelligence":2.8293414417857914,"speed":67.62138193768119},"rank":228,"open_board_rank":184,"api_board_rank":null}},{"key":"open-jev-json-canvas-joshuasp","name":"Open Jev JSON Canvas (JoshuaSP)","board":"open","source_url":"https://github.com/JoshuaSP/open-jev","axes":{"intelligence":61.19221807027989,"calibration":0,"speed":85.23407766636957,"cost":70.6750337639048},"capability":{"score":30.596109035139946,"rank":268,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":230,"api_board_rank":null},"composite_score":0,"rank":314,"open_board_rank":261,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009495085385709303,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 884,103 input + 0 output tokens over 1,624 decisions = 544.398399015 tokens/decision. GPU-time cost USD 0.0094950853857093/1,000; old list-price cost USD 0.0489958559113301/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0094950853857093/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15850554150529206,"p50_s_adjusted":0.46701108301058414,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15850554150529206,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.24580051028169692,"p95_s_adjusted":0.6416010205633939,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0,"capability":30.596109035139946,"axes":{"calibration":0,"cost":49.295219536461964,"intelligence":61.19221807027989,"speed":85.53332924416175},"rank":241,"open_board_rank":197,"api_board_rank":null}},{"key":"open-jev-zefan-27b-v1.1","name":"Open-Jev 27B v1.1 (Zefan Cai, LoRA + decision head on Qwen3.8-27B)","board":"open","source_url":"https://huggingface.co/ZefanCai/Open-Jev-27B-v1.1","axes":{"intelligence":52.45777412397041,"calibration":83.46836347105969,"speed":72.14822609431056,"cost":41.72719016000595},"capability":{"score":67.96306879751505,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.7× Jev (v1.5 reference)","latency 2.9× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":40.45143841684714,"rank":169,"open_board_rank":134,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0875844474931956,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,767,510 input + 0 output tokens over 1,624 decisions = 1704.1317734 tokens/decision. GPU-time cost USD 0.0875844474931956/1,000; old list-price cost USD 0.715735344827586/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0875844474931956/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.818691068328917,"p50_s_adjusted":1.787382136657834,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.818691068328917,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.6308073512278496,"p95_s_adjusted":3.411614702455699,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":2.8808671915746493,"capability":67.96306879751505,"axes":{"calibration":83.46836347105969,"cost":14.357426070787312,"intelligence":52.45777412397041,"speed":72.50871441156775},"rank":194,"open_board_rank":153,"api_board_rank":null}},{"key":"open-jev-zefan-2b","name":"Open-Jev 2B (Zefan Cai)","board":"open","source_url":"https://github.com/Zefan-Cai/Open-Jev","axes":{"intelligence":21.804441188912403,"calibration":73.61830158128652,"speed":77.13082993864059,"cost":69.61164724940987},"capability":{"score":47.71137138509946,"rank":224,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":188,"api_board_rank":null},"composite_score":8.766158492730982,"rank":250,"open_board_rank":203,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0103025585586246,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,767,510 input + 0 output tokens over 1,624 decisions = 1704.1317734 tokens/decision. GPU-time cost USD 0.0103025585586246/1,000; old list-price cost USD 0.170413177339901/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0103025585586246/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.45112593495287,"p50_s_adjusted":1.05225186990574,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.45112593495287,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.8449565254268235,"p95_s_adjusted":1.839913050853647,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.2383990318693976,"capability":47.71137138509946,"axes":{"calibration":73.61830158128652,"cost":33.05490478272432,"intelligence":21.804441188912403,"speed":77.13731150142542},"rank":192,"open_board_rank":152,"api_board_rank":null}},{"key":"open-jev-zefan-9b","name":"Open-Jev 9B (Zefan Cai)","board":"open","source_url":"https://github.com/Zefan-Cai/Open-Jev","axes":{"intelligence":43.88591848753404,"calibration":79.80507069734225,"speed":74.31878072613569,"cost":52.71772037731421},"capability":{"score":61.84549459243814,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 2.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":45.49003604451487,"rank":150,"open_board_rank":118,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03767697156750158,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,767,510 input + 0 output tokens over 1,624 decisions = 1704.1317734 tokens/decision. GPU-time cost USD 0.0376769715675016/1,000; old list-price cost USD 0.170413177339901/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0376769715675016/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.6436806479468942,"p50_s_adjusted":1.4373612958937882,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.6436806479468942,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.211844247172121,"p95_s_adjusted":2.5736884943442417,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.045611746249723,"capability":61.84549459243814,"axes":{"calibration":79.80507069734225,"cost":33.05490478272432,"intelligence":43.88591848753404,"speed":74.39486349899005},"rank":147,"open_board_rank":111,"api_board_rank":null}},{"key":"openai-decisions","name":"OpenAI Decisions (gpt-6-luna)","board":"api","source_url":null,"axes":{"intelligence":56.9,"calibration":90.1,"speed":88.31,"cost":48.6},"capability":{"score":73.5,"rank":75,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":21},"composite_score":62.37,"rank":126,"open_board_rank":null,"api_board_rank":28,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.0517,"basis":"measured: token usage of the 6 Oct 2026 A5 u P run x tariff USD 0.1/M input, 0.0/M output (spec openai-decisions.json; OpenAI Decisions API (POST /v1/decisions, gpt-6-luna) list price USD 0.10 per 1M input tokens, no output/cache charges; source https://developers.openai.com/api/docs/guides/decisions (checked 2026-10-06); actual usage), per 1,000 decisions on the v1.6.1 common cost basis (592 of 600 items; the 8 A5 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a5.json). List price as of 6 Oct 2026 (launch day); re-checked at each revision."},"latency":{"p50_s_raw":0.305,"p50_s_adjusted":0.305,"adjustment":"none (API)","p50_s_measured":0.305,"latency_basis":"api-end-to-end","p95_s_raw":0.484,"p95_s_adjusted":0.484,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":62.5,"capability":73.5,"axes":{"calibration":90.1,"cost":48.6,"intelligence":56.9,"speed":89.25},"rank":53,"open_board_rank":null,"api_board_rank":22}},{"key":"opendecision","name":"OpenDecision (ModernBERT-large zero-shot)","board":"open","source_url":"https://github.com/deepanwadhwa/OpenDecision","axes":{"intelligence":7.425225288997898,"calibration":68.56606448002526,"speed":86.36099014084385,"cost":79.10704862374335},"capability":{"score":37.99564488451158,"rank":255,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":218,"api_board_rank":null},"composite_score":0.5084959848229078,"rank":279,"open_board_rank":227,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004970862068965518,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.09270225139334798,"p50_s_adjusted":0.335404502786696,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09270225139334798,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2695895489188842,"p95_s_adjusted":0.6891790978377684,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.508795165212109,"capability":37.99564488451158,"axes":{"calibration":68.56606448002526,"cost":79.10704862374335,"intelligence":7.425225288997898,"speed":87.12856375997052},"rank":205,"open_board_rank":162,"api_board_rank":null}},{"key":"openjev-sglang","name":"openjev-sglang (Qwen3.6-35B-A3B on SGLang)","board":"api","source_url":"https://github.com/ekzhang/openjev-sglang","axes":{"intelligence":38.77,"calibration":78.57,"speed":77.66,"cost":35.63},"capability":{"score":58.67,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 4.3× Jev (v1.5 reference)","latency 2.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":15.38,"rank":234,"open_board_rank":null,"api_board_rank":47,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.13981028325123152,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: exact base-model market reference (operator tariff excluded by 30-day rule)"},"latency":{"p50_s_raw":0.635,"p50_s_adjusted":1.27,"adjustment":"x2 (demo)","p50_s_measured":0.635,"latency_basis":"demo-endpoint-x2","p95_s_raw":0.675,"p95_s_adjusted":1.35,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.38,"capability":58.67,"axes":{"calibration":78.57,"cost":35.63,"intelligence":38.77,"speed":77.71},"rank":160,"open_board_rank":null,"api_board_rank":39}},{"key":"openjev-thinking","name":"OpenJev (thinking, BF16)","board":"open","source_url":"https://github.com/razorback16/openjev","axes":{"intelligence":70.21463489688583,"calibration":81.33711087001693,"speed":72.91370874195911,"cost":54.86348744521688},"capability":{"score":75.77587288345137,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":68.40125787286172,"rank":112,"open_board_rank":88,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03195585327946334,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,383,393 input + 592,069 output tokens over 1,624 decisions = 1832.18103448 tokens/decision. GPU-time cost USD 0.0319558532794633/1,000; old list-price cost USD 0.241456939655172/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0319558532794633/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.8890150445513427,"p50_s_adjusted":1.9280300891026854,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.8890150445513427,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.250822310359217,"p95_s_adjusted":2.651644620718434,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":17.265218966202397,"capability":75.87983429229429,"axes":{"calibration":81.41345474033434,"cost":28.51480924396573,"intelligence":70.34621384425424,"speed":72.55270827201397},"rank":146,"open_board_rank":110,"api_board_rank":null}},{"key":"openjev-verdict","name":"openJev Verdict (heman10x, ModernBERT-base 151M)","board":"open","source_url":"https://github.com/Heman10x-NGU/openJev-verdict-2.0","axes":{"intelligence":4.1496399816416485,"calibration":49.11476555745841,"speed":78.45185691662326,"cost":86.62313527343974},"capability":{"score":26.63220276955003,"rank":273,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":234,"api_board_rank":null},"composite_score":0.09645563919820467,"rank":295,"open_board_rank":242,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.002791871921182266,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: v1.5 measured-input proxy; no exact base-model floor applies"},"latency":{"p50_s_raw":0.3206168063916266,"p50_s_adjusted":0.7912336127832532,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.3206168063916266,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.8275672776158899,"p95_s_adjusted":1.8051345552317797,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.09650709186186587,"capability":26.63220276955003,"axes":{"calibration":49.11476555745841,"cost":86.62313527343974,"intelligence":4.1496399816416485,"speed":79.40046783928},"rank":221,"open_board_rank":177,"api_board_rank":null}},{"key":"openjev-verdict-1.4","name":"openJev Verdict 1.4","board":"open","source_url":"https://huggingface.co/heman10x/rlcd-modernbert-151m","axes":{"intelligence":5.035901218168153,"calibration":75.69890703057408,"speed":80.93559275082234,"cost":86.62313527343974},"capability":{"score":40.36740412437111,"rank":251,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":214,"api_board_rank":null},"composite_score":0.17216463673473972,"rank":289,"open_board_rank":236,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.002791871921182266,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: v1.5 measured-input proxy; no exact base-model floor applies"},"latency":{"p50_s_raw":0.3353864559903741,"p50_s_adjusted":0.8207729119807482,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.3353864559903741,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.4161201062146574,"p95_s_adjusted":0.9822402124293148,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.17222636557331997,"capability":40.36740412437111,"axes":{"calibration":75.69890703057408,"cost":86.62313527343974,"intelligence":5.035901218168153,"speed":81.49274954261848},"rank":216,"open_board_rank":172,"api_board_rank":null}},{"key":"opensourcejev-qwen35-4b-q4km","name":"OpenSourceJev (Qwen3.5-4B Q4_K_M, native llama.cpp)","board":"open","source_url":"https://github.com/sabeel111/OpenSourceJev","axes":{"intelligence":23.217995107845883,"calibration":73.92484912653177,"speed":75.84507183307446,"cost":83.8256538982382},"capability":{"score":48.571422117188824,"rank":223,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":187,"api_board_rank":null},"composite_score":10.555669166289977,"rank":245,"open_board_rank":198,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003460547940236125,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 570,303 input + 1,624 output tokens over 1,624 decisions = 352.17179803 tokens/decision. GPU-time cost USD 0.00346054794023613/1,000; old list-price cost USD 0.0106851539408867/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00346054794023613/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.3206673553213477,"p50_s_adjusted":0.7913347106426954,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.3206673553213477,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.5697602639440447,"p95_s_adjusted":3.2895205278880892,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":10.263493713193125,"capability":48.571422117188824,"axes":{"calibration":73.92484912653177,"cost":69.13657649738097,"intelligence":23.217995107845883,"speed":77.06362388237805},"rank":171,"open_board_rank":132,"api_board_rank":null}},{"key":"phillylabs_philly_decider_v2","name":"Philly Decider v2 (PhillyLabs)","board":"open","source_url":"https://huggingface.co/PhillyLabs/philly-decider-v2","axes":{"intelligence":56.08237887792767,"calibration":82.5161952575238,"speed":93.1956789542462,"cost":79.09341092589379},"capability":{"score":69.29928706772574,"rank":111,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":84,"api_board_rank":null},"composite_score":75.01434382924442,"rank":29,"open_board_rank":19,"api_board_rank":null,"measured_in":"v1.7.10","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.10","measured_on":"2026-10-26","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004976067953430363,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 752,008 input + 0 output tokens over 1,485 decisions = 506.402693603 tokens/decision. GPU-time cost USD 0.00497606795343036/1,000; old list-price cost USD 0.0151920808080808/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00497606795343036/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02382839099999501,"p50_s_adjusted":0.19765678199999,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02382839099999501,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04619658659999573,"p95_s_adjusted":0.24239317319999146,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.10","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"phocinae-largha-150m-v1","name":"Phocinae-Largha-150M-v1","board":"open","source_url":"https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1","axes":{"intelligence":3.609767391262904,"calibration":56.87058116402116,"speed":76.50701878654316,"cost":95.30667675976518},"capability":{"score":30.240174277642033,"rank":269,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":231,"api_board_rank":null},"composite_score":0.06552619942227712,"rank":299,"open_board_rank":246,"api_board_rank":null,"measured_in":"v1.6.15","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0014336531986531987,"basis":"ESTIMATE (*est.), size-class reference; self-hosted open weights, no public bookable tariff. A 144.3M ModernBERT encoder with no decoder lineage, so the base-size-encoder line is used unchanged - USD 0.005 per 1M input and 0.0 per 1M output, the line the PUBLISHED v1.6 field already carries on modernbert-jev-v2, tacet-sonata and watt-flash-0.1 (verified in their own registry entries; modernbert-jev-v2 was measured in v1.6.8 on the earlier v1.6-regular-20261010-a2 draw, while tacet-sonata and watt-flash-0.1 are v1.6.1 rows - none of the three was measured on THIS draw, which carries no ModernBERT row), backed by the dated 25 Sep 2026 DeepInfra snapshot sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f in which thenlper/gte-base, BAAI/bge-base-en-v1.5, intfloat/e5-base-v2 and sentence-transformers/all-mpnet-base-v2 all price at USD 0.005 per 1M input and 0 output - and The frozen pricing reference is unchanged. Output tokens are a literal 0 reported by the author's own server (an encoder read-out; nothing is generated). x this row's own measured input tokens, which are the author's server's own count. Estimated, not charged. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.001434 per 1,000 decisions (0.022x the official cost cap USD 0.064595), against USD 0.001440 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.484241258000111,"p50_s_adjusted":1.1184825160002219,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.484241258000111,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.9241689420050533,"p95_s_adjusted":1.9983378840101065,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.06554752053348605,"capability":30.240174277642033,"axes":{"calibration":56.87058116402116,"cost":95.30667675976518,"intelligence":3.609767391262904,"speed":77.11764018052709},"rank":225,"open_board_rank":181,"api_board_rank":null}},{"key":"piccolo-27b","name":"Piccolo 27B (agentscope-ai)","board":"open","source_url":"https://huggingface.co/agentscope-ai/Piccolo-27B","axes":{"intelligence":75.75043982824246,"calibration":91.42754536732097,"speed":81.6188831463648,"cost":53.102986669932804},"capability":{"score":83.58899259778173,"rank":13,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":8,"api_board_rank":null},"composite_score":72.43455094447503,"rank":69,"open_board_rank":53,"api_board_rank":null,"measured_in":"v1.8.3","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.3","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0365791641698085,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,059,753 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0365791641698085/1,000; old list-price estimate USD 0.138785651443922/1,000 (qwen/qwen3.5-27b at USD 0.195/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0365791641698085/1,000. Estimated, not a bill. USD 0.0365791642 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.036480179 over all 1,500 items)."},"latency":{"p50_s_raw":0.22704334299999118,"p50_s_adjusted":0.6040866859999824,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.22704334299999118,"latency_basis":"api-end-to-end","p95_s_raw":0.49514120025007174,"p95_s_adjusted":1.1402824005001435,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"piccolo-4b","name":"Piccolo 4B (agentscope-ai)","board":"open","source_url":"https://huggingface.co/agentscope-ai/Piccolo-4B","axes":{"intelligence":56.34925103899451,"calibration":86.08773956719725,"speed":88.05647378591166,"cost":74.46784937746082},"capability":{"score":71.21849530309588,"rank":92,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":69,"api_board_rank":null},"composite_score":73.87231626519474,"rank":46,"open_board_rank":33,"api_board_rank":null,"measured_in":"v1.8.2","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.2","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.007096949050642927,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,075,415 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00709694905064293/1,000; old list-price estimate USD 0.0216671927468099/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00709694905064293/1,000. Estimated, not a bill. Reference deepinfra:Qwen/Qwen3.5-4B is frozen at the 25 Sep cut-off and was deprecated afterwards (replaced by Qwen/Qwen3.5-9B); re-score if it changes (I-3). USD 0.00709694905 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00707726606 over all 1,500 items)."},"latency":{"p50_s_raw":0.0937403720000134,"p50_s_adjusted":0.3374807440000268,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.0937403720000134,"latency_basis":"api-end-to-end","p95_s_raw":0.1567787561499813,"p95_s_adjusted":0.46355751229996256,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.2","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"piccolo-9b","name":"Piccolo 9B (agentscope-ai)","board":"open","source_url":"https://huggingface.co/agentscope-ai/Piccolo-9B","axes":{"intelligence":60.88078730703727,"calibration":87.64362592882337,"speed":87.51139650157955,"cost":63.90237383411994},"capability":{"score":74.26220661793032,"rank":69,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":50,"api_board_rank":null},"composite_score":72.84469830279757,"rank":63,"open_board_rank":48,"api_board_rank":null,"measured_in":"v1.8.2","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.2","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.015968135363946586,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,075,415 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0159681353639466/1,000; old list-price estimate USD 0.0722239758226998/1,000 (qwen/qwen3.5-9b at USD 0.1/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0159681353639466/1,000. Estimated, not a bill. USD 0.0159681354 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0159238486 over all 1,500 items)."},"latency":{"p50_s_raw":0.09442659100000128,"p50_s_adjusted":0.3388531820000026,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.09442659100000128,"latency_basis":"api-end-to-end","p95_s_raw":0.1867090770999425,"p95_s_adjusted":0.523418154199885,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.2","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"pjev-2b","name":"pjev-2b (Kailune AI, frozen MiniCPM5-2B + merged LoRA r64, letter-logit readout, T=1 stub)","board":"open","source_url":"https://huggingface.co/Kailune-AI/pjev-2b","axes":{"intelligence":28.75448619260912,"calibration":84.15501405892886,"speed":92.67163234372039,"cost":63.52461077797332},"capability":{"score":56.45475012576899,"rank":192,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":158,"api_board_rank":null},"composite_score":18.074372740962183,"rank":221,"open_board_rank":178,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01643789898989899,"basis":"ESTIMATE (*est.), size-class reference, carried over unchanged from PREREG-R36.md section 3.2. The UNMODIFIED frozen pricing_v15 path is asserted: openbmb/MiniCPM5-2B is in UNLISTED_BASE_MODELS - explicitly reviewed against both 25 Sep 2026 catalogs and not sold - so reference() returns no market reference rather than raising, and no exact base-model floor applies. The board already prices a row on THE SAME BASE this way: decision-2b (FlyMy.AI, openbmb/MiniCPM5-2B + LoRA and pointer head) carries the DeepInfra 2-4B size-class reference list price of USD 0.03 per 1M input and USD 0.0 per 1M output from the frozen 25 Sep 2026 snapshot (sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f). pjev-2b is the same shape: 2.52B, one forward pass with logits_to_keep=1, NOTHING GENERATED - the author's serve.py reports usage.output_tokens as a literal 0, not a decision count - so the output term is 0.0, not 0.15. Multiplied by this row's own measured input tokens. The frozen pricing reference is unchanged. Free weights are never cost 100 and never cost 0. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.016438 per 1,000 decisions (0.254x the official cost cap USD 0.064595), against USD 0.016274 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.035382037499971375,"p50_s_adjusted":0.22076407499994274,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.035382037499971375,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04742732714998965,"p95_s_adjusted":0.2448546542999793,"n":1500},"method":"v1.7","rescored_in":null,"remeasured_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.759681433648748,"capability":57.71754900661349,"axes":{"calibration":86.466579728622,"cost":63.52461077797332,"intelligence":28.96851828460498,"speed":49.10963929819891},"rank":156,"open_board_rank":118,"api_board_rank":null}},{"key":"plumb-4b","name":"Plumb-4B (crh225, JevK5 v0.2 + LoRA)","board":"open","source_url":"https://huggingface.co/crh225/plumb-4b","axes":{"intelligence":42.83684414340664,"calibration":87.80311208802716,"speed":93.68902806065091,"cost":77.61552273233744},"capability":{"score":65.3199781157169,"rank":131,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":103,"api_board_rank":null},"composite_score":50.367731512900676,"rank":143,"open_board_rank":112,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.021091305068694055,"p50_s_adjusted":0.1921826101373881,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.021091305068694055,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03626360079506411,"p95_s_adjusted":0.22252720159012823,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":48.28224126395837,"capability":65.42184422831923,"axes":{"calibration":87.88101858186977,"cost":63.07359395281681,"intelligence":42.9626698747687,"speed":93.83356539601934},"rank":67,"open_board_rank":43,"api_board_rank":null}},{"key":"pplx-decider-flash-v1.2","name":"pplx-decider-flash-v1.2 (Perplexity, Qwen3.5-4B full fine-tune, marker readout, self-hosted)","board":"open","source_url":"https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2","axes":{"intelligence":45.47596631214893,"calibration":77.99716887881635,"speed":91.46033110721706,"cost":78.96751116224428},"capability":{"score":61.73656759548264,"rank":159,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":127,"api_board_rank":null},"composite_score":56.65175338549098,"rank":135,"open_board_rank":105,"api_board_rank":null,"measured_in":"v1.6.31","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-19","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005024385588609707,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 759,310 input + 0 output tokens over 1,485 decisions = 511.31986532 tokens/decision. GPU-time cost USD 0.00502438558860971/1,000; old list-price cost USD 0.015339595959596/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00502438558860971/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0443364574999805,"p50_s_adjusted":0.238672914999961,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0443364574999805,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07466965455015497,"p95_s_adjusted":0.2993393091003099,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":56.60442934796988,"capability":62.51340446094119,"axes":{"calibration":78.68376661171601,"cost":64.42558238272366,"intelligence":46.343042310166375,"speed":91.76093007933393},"rank":59,"open_board_rank":35,"api_board_rank":null}},{"key":"pplx-decider-v1-1-27b","name":"Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout)","board":"open","source_url":"https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b","axes":{"intelligence":75.59095372376527,"calibration":89.82871864473867,"speed":86.35967790128073,"cost":57.33119184627563},"capability":{"score":82.70983618425197,"rank":14,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":9,"api_board_rank":null},"composite_score":74.92968283438655,"rank":30,"open_board_rank":20,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.026441957638144272,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11830.7366892 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=trunk-equivalent). The row's own scored cost basis: 764,265 input + 0 output tokens over 1,477 decisions = 517.444143534 tokens/decision. GPU-time cost USD 0.0264419576381443/1,000; old list-price cost USD 0.21732654028436/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0264419576381443/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13107524551742245,"p50_s_adjusted":0.4121504910348449,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13107524551742245,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2055087369008106,"p95_s_adjusted":0.5610174738016211,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.617642759667973,"capability":82.79186839726479,"axes":{"calibration":89.86695733940225,"cost":29.886617009940878,"intelligence":75.71677945512734,"speed":86.79279341461623},"rank":127,"open_board_rank":93,"api_board_rank":null}},{"key":"pplx-decider-v1.2","name":"pplx-decider-v1.2 (Perplexity, Qwen3.8-27B full fine-tune, marker readout, self-hosted)","board":"open","source_url":"https://huggingface.co/perplexity-ai/pplx-decider-v1.2","axes":{"intelligence":71.71777888018025,"calibration":89.96214607141948,"speed":85.96624125269479,"cost":57.41150540159515},"capability":{"score":80.83996247579987,"rank":32,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":24,"api_board_rank":null},"composite_score":73.92360529540504,"rank":45,"open_board_rank":32,"api_board_rank":null,"measured_in":"v1.6.33","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-20","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02627946300596891,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 759,310 input + 0 output tokens over 1,485 decisions = 511.31986532 tokens/decision. GPU-time cost USD 0.0262794630059689/1,000; old list-price cost USD 0.214754343434343/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0262794630059689/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.14681606799996416,"p50_s_adjusted":0.44363213599992835,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.14681606799996416,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2103138675499906,"p95_s_adjusted":0.5706277350999812,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.73512390064423,"capability":81.74185907802455,"axes":{"calibration":90.77013189733844,"cost":30.0417413123765,"intelligence":72.71358625871065,"speed":86.32357007699105},"rank":125,"open_board_rank":91,"api_board_rank":null}},{"key":"purplemist-flash-1.0","name":"PurpleMIST-Flash-1.0 (Olaverse)","board":"open","source_url":"https://huggingface.co/olaverse/PurpleMIST-Flash-1.0","axes":{"intelligence":36.57669575617585,"calibration":83.64479179791499,"speed":87.71257539670552,"cost":null},"capability":{"score":60.11074377704542,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.28","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-17","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base and raises rather than estimating; one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 6202ef78..., the module in force at scoring time; the later re-freezes only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base Qwen/Qwen3.5-9B-Base (the module RAISES 'base model lacks an explicit market-listing decision', for every spelling tried); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void - the same treatment as the live jevora-9b-v1 and kev-9b-v2 rows on that same declared base. The tokens needed for a later reprice are measured and on file: the author's server reports NO usage block at all, so the input tokens were counted OFFLINE with the author's own pinned tokenizer and their own unchanged encode() on the exact prompts their server builds from the runner's request bodies (the pricing update, as accepted for the sibling row): 700,492 input tokens over the 1,485 answered items, 0 output (nothing is generated). That count is byte-for-byte identical, item by item, to the count accepted for purplemist-mini-1.0 - same rendered-prompt-ids sha256 ad0810696b7e80eba87192caa70f57c19cdbee103e37f3d9cd163ef5df434c86 - which it must be, because the rendered prompt depends only on the tokenizer, encode(), the two flags and the request body, all byte-identical between the two rows, and not on the weights."},"latency":{"p50_s_raw":0.09043012899928726,"p50_s_adjusted":0.33086025799857455,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09043012899928726,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.18089857094928447,"p95_s_adjusted":0.511797141898569,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":60.11074377704542,"axes":{"calibration":83.64479179791499,"cost":null,"intelligence":36.57669575617585,"speed":88.25066979413353},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"purplemist-mini-1.0","name":"PurpleMIST-Mini-1.0 (Olaverse)","board":"open","source_url":"https://huggingface.co/olaverse/PurpleMIST-Mini-1.0","axes":{"intelligence":8.377579633895749,"calibration":77.68611335416006,"speed":88.54013979409211,"cost":86.34644068316165},"capability":{"score":43.0318464940279,"rank":241,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":204,"api_board_rank":null},"composite_score":0.7239454481982587,"rank":276,"open_board_rank":224,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0028517972369194594,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 700,492 input + 0 output tokens over 1,485 decisions = 471.711784512 tokens/decision. GPU-time cost USD 0.00285179723691946/1,000; old list-price cost USD 0.00943423569023569/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00285179723691946/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07631494899999325,"p50_s_adjusted":0.3026298979999865,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07631494899999325,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.15623000759997013,"p95_s_adjusted":0.4624600151999403,"n":1500},"method":"v1.7","rescored_in":null,"remeasured_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.6877481478120564,"capability":43.033098052325656,"axes":{"calibration":77.68645539289165,"cost":70.75879834655666,"intelligence":8.37974071175967,"speed":58.74735259439709},"rank":202,"open_board_rank":159,"api_board_rank":null}},{"key":"quyet-1-0-large","name":"Quyet-1.0-Large (Chinh Nguyen, Gemma-4-31B decoder, option-letter logits)","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.0-Large","axes":{"intelligence":73.40135940607429,"calibration":89.97624968733336,"speed":86.35482296351647,"cost":55.868159288206435},"capability":{"score":81.68880454670384,"rank":23,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":17,"api_board_rank":null},"composite_score":73.77734958545854,"rank":47,"open_board_rank":34,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029584304036350536,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 803,169 input + 0 output tokens over 1,624 decisions = 494.562192118 tokens/decision. GPU-time cost USD 0.0295843040363505/1,000; old list-price cost USD 0.0445105972906404/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0295843040363505/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12244235770776868,"p50_s_adjusted":0.3948847154155374,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12244235770776868,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.21810105652548373,"p95_s_adjusted":0.5862021130509675,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":71.38706329141472,"capability":81.68880454670384,"axes":{"calibration":89.97624968733336,"cost":50.546097335562465,"intelligence":73.40135940607429,"speed":86.87440406834797},"rank":19,"open_board_rank":7,"api_board_rank":null}},{"key":"quyet-1-0-medium","name":"Quyet-1.0-Medium (Chinh Nguyen, Qwen3.5-4B decoder, option-letter logits)","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.0-Medium","axes":{"intelligence":41.73350604436517,"calibration":76.76007761870373,"speed":89.14092560273156,"cost":77.97650072588999},"capability":{"score":59.24679183153445,"rank":172,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":140,"api_board_rank":null},"composite_score":45.659799874048836,"rank":149,"open_board_rank":117,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005421464801247415,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 896,009 input + 0 output tokens over 1,624 decisions = 551.729679803 tokens/decision. GPU-time cost USD 0.00542146480124741/1,000; old list-price cost USD 0.0165518903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00542146480124741/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0816944474936463,"p50_s_adjusted":0.3133888949872926,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0816944474936463,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11944368409385787,"p95_s_adjusted":0.38888736818771574,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":43.58925238992216,"capability":59.24679183153445,"axes":{"calibration":76.76007761870373,"cost":63.43457194636937,"intelligence":41.73350604436517,"speed":89.46638461924755},"rank":73,"open_board_rank":48,"api_board_rank":null}},{"key":"quyet-1-0-small","name":"Quyet-1.0-Small (Chinh Nguyen, SEA-LION-ModernBERT-300M encoder with fixed heads)","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.0-Small","axes":{"intelligence":4.019045994532796,"calibration":60.96568104722222,"speed":94.90280801143004,"cost":79.56261245432609},"capability":{"score":32.492363520877504,"rank":266,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":228,"api_board_rank":null},"composite_score":0.08963662915567427,"rank":296,"open_board_rank":243,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004800055418719211,"basis":"v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (v1.5-M2 base-model reference; self-hosted open weights, no public tariff): DeepInfra same-size encoders (bge-large, e5-large, Qwen3-Embedding-0.6B) $0.01/M input (SEA-LION-ModernBERT-300M has no market listing; the ModernBERT-large-class reference of the Laya/Certo/OpenDecision rows) x the package's own measured usage.input_tokens; one forward pass, nothing generated; run in process on 1x H100 80GB (Lium), quyet 1.0.0; estimated, not charged"},"latency":{"p50_s_raw":0.014559379997081123,"p50_s_adjusted":0.17911875999416224,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.014559379997081123,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.015270974154555006,"p95_s_adjusted":0.18054194830911,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.08963716950381444,"capability":32.492363520877504,"axes":{"calibration":60.96568104722222,"cost":79.56261245432609,"intelligence":4.019045994532796,"speed":94.91846454857637},"rank":222,"open_board_rank":178,"api_board_rank":null}},{"key":"quyet-1-0-small-en","name":"Quyet-1.0-Small-EN (Chinh Nguyen, ModernBERT-base encoder with fixed heads)","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.0-Small-EN","axes":{"intelligence":6.054813340422714,"calibration":73.02134054125926,"speed":94.88950688832834,"cost":89.11296503596958},"capability":{"score":39.53807694084099,"rank":252,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":215,"api_board_rank":null},"composite_score":0.2923902780429571,"rank":284,"open_board_rank":231,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0023062222906403943,"basis":"v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (v1.5-M2 base-model reference; self-hosted open weights, no public tariff): DeepInfra base-size encoders (bge-base, e5-base, gte-base, all-mpnet-base) $0.005/M input (ModernBERT-base has no market listing; base-size encoder reference) x the package's own measured usage.input_tokens; one forward pass, nothing generated; run in process on 1x H100 80GB (Lium), quyet 1.0.0; estimated, not charged"},"latency":{"p50_s_raw":0.01468971849681111,"p50_s_adjusted":0.17937943699362222,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.01468971849681111,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.01541628500490333,"p95_s_adjusted":0.18083257000980665,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.29239342204088814,"capability":39.53807694084099,"axes":{"calibration":73.02134054125926,"cost":89.11296503596958,"intelligence":6.054813340422714,"speed":94.90893356488051},"rank":211,"open_board_rank":167,"api_board_rank":null}},{"key":"quyet-1-0-tiny","name":"Quyet-1.0-Tiny (Chinh Nguyen, mmBERT-small encoder (16 layers kept) with fixed heads)","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.0-Tiny","axes":{"intelligence":3.6964771541368573,"calibration":58.46458026762962,"speed":95.07796177344144,"cost":88.43590631389968},"capability":{"score":31.08052871088324,"rank":267,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":229,"api_board_rank":null},"composite_score":0.07064720497480408,"rank":298,"open_board_rank":245,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0024292364532019703,"basis":"v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (v1.5-M2 base-model reference; self-hosted open weights, no public tariff): DeepInfra base-size encoders (bge-base, e5-base, gte-base, all-mpnet-base) $0.005/M input (mmBERT-small 16-layer, 183M; base-size class reference, errs high vs the small-encoder $0.004 class) x the package's own measured usage.input_tokens; one forward pass, nothing generated; run in process on 1x H100 80GB (Lium), quyet 1.0.0; estimated, not charged"},"latency":{"p50_s_raw":0.012781936500687152,"p50_s_adjusted":0.1755638730013743,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.012781936500687152,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.013458317358163184,"p95_s_adjusted":0.17691663471632635,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.07064766318734213,"capability":31.08052871088324,"axes":{"calibration":58.46458026762962,"cost":88.43590631389968,"intelligence":3.6964771541368573,"speed":95.09610910731021},"rank":224,"open_board_rank":180,"api_board_rank":null}},{"key":"quyet-1-2-large","name":"Quyet-1.2-Large","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.2-Large","axes":{"intelligence":74.51696403947234,"calibration":88.9668822338541,"speed":88.26322277501028,"cost":61.7344191165331},"capability":{"score":81.74192313666322,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":76.6458737791496,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.14","measured_on":"2026-11-16","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.01885903049573297,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.018859030495733/1,000; old list-price estimate USD 0.0283740564137005/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.018859030495733/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.07942035999997188,"p50_s_adjusted":0.3088407199999438,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07942035999997188,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16649779545002158,"p95_s_adjusted":0.4829955909000432,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.14","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"quyet-1-2-small","name":"Quyet-1.2-Small","board":"open","source_url":"https://huggingface.co/chinhnc/Quyet-1.2-Small","axes":{"intelligence":31.79407821885479,"calibration":81.39847698823577,"speed":93.3863994106336,"cost":88.8673177026433},"capability":{"score":56.59627760354528,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":24.618214654792457,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.14","measured_on":"2026-11-16","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.0023501165987887835,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: small gemma, 4B active parameters; throughput 80585.75 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.00235011659878878/1,000; old list-price estimate USD 0.00626521155137676/1,000 (deepinfra:google/gemma-4-E4B-it at USD 0.02/M input and USD 0.1/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00235011659878878/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.023549430999992182,"p50_s_adjusted":0.19709886199998436,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.023549430999992182,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.04131773760000214,"p95_s_adjusted":0.23263547520000427,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.14","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"qwen3-reranker-4b","name":"Qwen3-Reranker-4B","board":"open","source_url":"https://huggingface.co/Qwen/Qwen3-Reranker-4B","axes":{"intelligence":16.337944014530347,"calibration":73.15350658389404,"speed":81.05912758410997,"cost":60.56419192241651},"capability":{"score":44.7457252992122,"rank":233,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":196,"api_board_rank":null},"composite_score":4.117471541314488,"rank":261,"open_board_rank":212,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020631316234009007,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 3,409,751 input + 0 output tokens over 1,624 decisions = 2099.60036946 tokens/decision. GPU-time cost USD 0.020631316234009/1,000; old list-price cost USD 0.0524900092364532/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.020631316234009/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.21138389385305345,"p50_s_adjusted":0.5727677877061069,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.21138389385305345,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.6090355303371323,"p95_s_adjusted":1.3680710606742645,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.715006798009075,"capability":44.7457252992122,"axes":{"calibration":73.15350658389404,"cost":48.39770052432463,"intelligence":16.337944014530347,"speed":82.14971164581814},"rank":187,"open_board_rank":147,"api_board_rank":null}},{"key":"qwen3.8-27b","name":"Qwen3.8 27B (Chutes TEE)","board":"api","source_url":"https://benchmarkheaven.com/jev-models/qwen3.8-27b","axes":{"intelligence":33.36,"calibration":63.81,"speed":54.99,"cost":0},"capability":{"score":48.585,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 67.4× Jev (v1.5 reference)","latency 13.5× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0,"rank":315,"open_board_rank":null,"api_board_rank":54,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":2.1783887438423646,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: exact base-model market reference (operator tariff excluded by 30-day rule)"},"latency":{"p50_s_raw":8.35,"p50_s_adjusted":8.35,"adjustment":"none (API)","p50_s_measured":8.35,"latency_basis":"api-end-to-end","p95_s_raw":37.933,"p95_s_adjusted":37.933,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":233,"n_over_limit_answered_ok":225,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0,"capability":98.19,"axes":{"calibration":100,"cost":0,"intelligence":96.38,"speed":57.43},"rank":242,"open_board_rank":null,"api_board_rank":45}},{"key":"qwen35-9b-jev-data-mix-v2","name":"Qwen3.5-9B Jev-like data-mix v2","board":"open","source_url":"https://huggingface.co/jsaurabh/qwen3.5-9b-jev-data-mix-v2","axes":{"intelligence":41.94184606895763,"calibration":79.78895435560509,"speed":85.22302455389745,"cost":65.35303366044648},"capability":{"score":60.86540021228136,"rank":167,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":135,"api_board_rank":null},"composite_score":44.38643689711108,"rank":154,"open_board_rank":122,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.014285612906521182,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,049,330 input + 0 output tokens over 1,624 decisions = 646.139162562 tokens/decision. GPU-time cost USD 0.0142856129065212/1,000; old list-price cost USD 0.0646139162561576/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0142856129065212/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1711398959159851,"p50_s_adjusted":0.49227979183197024,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1711398959159851,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.23010936684440822,"p95_s_adjusted":0.6102187336888164,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":33.58966697347588,"capability":60.86540021228136,"axes":{"calibration":79.78895435560509,"cost":45.690218065856605,"intelligence":41.94184606895763,"speed":85.40793837616167},"rank":94,"open_board_rank":64,"api_board_rank":null}},{"key":"raw-phi-4-mini","name":"Raw Phi-4 mini direct logits","board":"open","source_url":"https://huggingface.co/microsoft/Phi-4-mini-instruct","axes":{"intelligence":13.660237978082595,"calibration":57.015465878675315,"speed":91.5252083185614,"cost":53.22051180195924},"capability":{"score":35.33785192837895,"rank":260,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":223,"api_board_rank":null},"composite_score":2.4785276781687653,"rank":267,"open_board_rank":217,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03625068965517241,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.043601186480373144,"p50_s_adjusted":0.23720237296074628,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.043601186480373144,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0733645498054102,"p95_s_adjusted":0.2967290996108204,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":2.4794266829120515,"capability":35.33785192837895,"axes":{"calibration":57.015465878675315,"cost":53.22051180195924,"intelligence":13.660237978082595,"speed":91.89255362835293},"rank":195,"open_board_rank":154,"api_board_rank":null}},{"key":"raw-qwen3-0.6b","name":"Raw Qwen3 0.6B direct logits","board":"open","source_url":"https://huggingface.co/Qwen/Qwen3-0.6B","axes":{"intelligence":5.554864880982027,"calibration":10.94062892769364,"speed":93.42809294775304,"cost":84.13747424220067},"capability":{"score":8.247746904337834,"rank":282,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":243,"api_board_rank":null},"composite_score":0.1679189307325968,"rank":290,"open_board_rank":237,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003378709544090625,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.6B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 907,601 input + 0 output tokens over 1,624 decisions = 558.867610837 tokens/decision. GPU-time cost USD 0.00337870954409063/1,000; old list-price cost USD 0.00558867610837438/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00337870954409063/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.029423376079648733,"p50_s_adjusted":0.20884675215929746,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.029423376079648733,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03372589310165495,"p95_s_adjusted":0.2174517862033099,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.1673536172919218,"capability":8.247746904337834,"axes":{"calibration":10.94062892769364,"cost":77.58073177551512,"intelligence":5.554864880982027,"speed":93.52703208725202},"rank":217,"open_board_rank":173,"api_board_rank":null}},{"key":"raw-qwen3-1.7b","name":"Raw Qwen3 1.7B direct logits","board":"open","source_url":"https://huggingface.co/Qwen/Qwen3-1.7B","axes":{"intelligence":5.124286554530066,"calibration":11.516326812547591,"speed":93.09539316451654,"cost":84.13747424220067},"capability":{"score":8.320306683538828,"rank":281,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":242,"api_board_rank":null},"composite_score":0.13792518704402346,"rank":292,"open_board_rank":239,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003378709544090625,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 1.7B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 907,601 input + 0 output tokens over 1,624 decisions = 558.867610837 tokens/decision. GPU-time cost USD 0.00337870954409063/1,000; old list-price cost USD 0.0111773522167488/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00337870954409063/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0324228135868907,"p50_s_adjusted":0.2148456271737814,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0324228135868907,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0391048670047894,"p95_s_adjusted":0.22820973400957878,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.13671939226506033,"capability":8.320306683538828,"axes":{"calibration":11.516326812547591,"cost":68.54983190559568,"intelligence":5.124286554530066,"speed":93.23530725162705},"rank":219,"open_board_rank":175,"api_board_rank":null}},{"key":"raw-qwen3-4b-instruct-2507","name":"Raw Qwen3 4B Instruct 2507 direct logits","board":"open","source_url":"https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507","axes":{"intelligence":33.95117742100547,"calibration":35.18175191334879,"speed":90.72731866330435,"cost":77.90260990054026},"capability":{"score":34.56646466717713,"rank":262,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":224,"api_board_rank":null},"composite_score":22.563881701127844,"rank":208,"open_board_rank":167,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005452299072585267,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 901,105 input + 0 output tokens over 1,624 decisions = 554.867610837 tokens/decision. GPU-time cost USD 0.00545229907258527/1,000; old list-price cost USD 0.0166460283251232/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00545229907258527/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.049780144123360515,"p50_s_adjusted":0.24956028824672102,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.049780144123360515,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09445822748821228,"p95_s_adjusted":0.33891645497642453,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.792436038282993,"capability":34.56646466717713,"axes":{"calibration":35.18175191334879,"cost":63.36068112101963,"intelligence":33.95117742100547,"speed":91.16296741568695},"rank":116,"open_board_rank":83,"api_board_rank":null}},{"key":"raw-qwen3-8b","name":"Raw Qwen3 8B direct logits","board":"open","source_url":"https://huggingface.co/Qwen/Qwen3-8B","axes":{"intelligence":26.54912920590792,"calibration":33.721548452619835,"speed":88.245839907826,"cost":68.77812304352645},"capability":{"score":30.13533882926388,"rank":270,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":232,"api_board_rank":null},"composite_score":12.101558108732878,"rank":242,"open_board_rank":195,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.010983208595174725,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 8B active parameters; throughput 30762.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 907,601 input + 0 output tokens over 1,624 decisions = 558.867610837 tokens/decision. GPU-time cost USD 0.0109832085951747/1,000; old list-price cost USD 0.0653875104679803/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0109832085951747/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.08167324960231781,"p50_s_adjusted":0.31334649920463564,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08167324960231781,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16397978526074436,"p95_s_adjusted":0.47795957052148874,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":9.30125972789427,"capability":30.13533882926388,"axes":{"calibration":33.721548452619835,"cost":45.53515592313027,"intelligence":26.54912920590792,"speed":88.65282990878164},"rank":174,"open_board_rank":135,"api_board_rank":null}},{"key":"reflex-4b","name":"reflex 4B (kshetrajna12)","board":"open","source_url":"https://github.com/kshetrajna12/reflex","axes":{"intelligence":30.80973256512164,"calibration":86.9025081232179,"speed":68.75878067126158,"cost":77.6846417557845},"capability":{"score":58.85612034416977,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 4.7× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":21.277263784731073,"rank":212,"open_board_rank":170,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005544281528016589,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 916,307 input + 0 output tokens over 1,624 decisions = 564.228448276 tokens/decision. GPU-time cost USD 0.00554428152801659/1,000; old list-price cost USD 0.0169268534482759/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00554428152801659/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":1.3620368089759722,"p50_s_adjusted":2.8740736179519444,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":1.3620368089759722,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":2.240229508350603,"p95_s_adjusted":4.630459016701207,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.637311182702398,"capability":58.95740703790636,"axes":{"calibration":86.97925577932901,"cost":63.142712976263866,"intelligence":30.935558296483705,"speed":68.7900017304639},"rank":126,"open_board_rank":92,"api_board_rank":null}},{"key":"reflex-instinct-0.6b","name":"Reflex Instinct 0.6B (Atlas AI)","board":"open","source_url":"https://huggingface.co/Atlas-AI-research/reflex-instinct-0.6b","axes":{"intelligence":3.96006223937058,"calibration":73.15265267444973,"speed":58.9717141639196,"cost":86.36009990882481},"capability":{"score":38.556357456910156,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 6.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.08513390654661894,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.15","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":false,"listing":"remeasure-pending","not_ranked_because":"Re-measure pending under the 10 s rule: this row's per-decision times were recorded on a loaded CPU host, so they cannot be scored under the time limit. Shown with its v1.6 values until it is measured again.","price":{"kind":"estimate","usd_per_1000_decisions":0.002848809024103469,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.6B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 699,758 input + 0 output tokens over 1,485 decisions = 471.217508418 tokens/decision. GPU-time cost USD 0.00284880902410347/1,000; old list-price cost USD 0.00471217508417508/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00284880902410347/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":1.8723777614941355,"p50_s_adjusted":3.894755522988271,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":1.8723777614941355,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":16.192409005090287,"p95_s_adjusted":32.53481801018057,"n":124},"method":"v1.6","rescored_in":null,"method_status":"v1.6-pending","decision_limit":null,"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.08485997925452016,"capability":38.556357456910156,"axes":{"calibration":73.15265267444973,"cost":79.80335744213926,"intelligence":3.96006223937058,"speed":58.9717141639196},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"rene-1-31b-fp8","name":"René-1 31B FP8 (salfatigroup, Gemma 4 31B full fine-tune, one-pass option readout)","board":"open","source_url":"https://huggingface.co/salfatigroup/rene-1-31b-fp8","axes":{"intelligence":61.72139577013068,"calibration":90.18677086647014,"speed":86.74281071257914,"cost":51.285700298534536},"capability":{"score":75.9540833183004,"rank":61,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":46,"api_board_rank":null},"composite_score":68.59077119594947,"rank":110,"open_board_rank":86,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.042054259482183,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,038,365 input + 0 output tokens over 1,477 decisions = 703.023019634 tokens/decision. GPU-time cost USD 0.042054259482183/1,000; old list-price cost USD 0.0632720717670955/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.042054259482183/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11330710025504231,"p50_s_adjusted":0.37661420051008465,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11330710025504231,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.20605559917166827,"p95_s_adjusted":0.5621111983433366,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":55.84876743320955,"capability":75.9540833183004,"axes":{"calibration":90.18677086647014,"cost":45.963638345890566,"intelligence":61.72139577013068,"speed":87.11720932683018},"rank":60,"open_board_rank":36,"api_board_rank":null}},{"key":"respan-span-02","name":"Respan Span-02","board":"open","source_url":"https://huggingface.co/lihe07/span-02","axes":{"intelligence":60.897967232257315,"calibration":77.78014902808363,"speed":85.7109495365315,"cost":53.528574234393844},"capability":{"score":69.33905813017047,"rank":110,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":83,"api_board_rank":null},"composite_score":67.0843877356023,"rank":120,"open_board_rank":94,"api_board_rank":null,"measured_in":"v1.6.18","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-14","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03540360816823925,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 776,560 input + 102,328 output tokens over 1,485 decisions = 591.843771044 tokens/decision. GPU-time cost USD 0.0354036081682393/1,000; old list-price cost USD 0.0677365656565656/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0354036081682393/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.030947174491302576,"p50_s_adjusted":0.21189434898260515,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.030947174491302576,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.5585132111977145,"p95_s_adjusted":1.2670264223954288,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":51.45147163621532,"capability":69.33905813017047,"axes":{"calibration":77.78014902808363,"cost":45.07530479505206,"intelligence":60.897967232257315,"speed":85.32322293910049},"rank":64,"open_board_rank":40,"api_board_rank":null}},{"key":"roudone-decider-4b","name":"Decider-4B (roudone, NeoHorse 4B pointer model, author server, self-hosted)","board":"open","source_url":"https://huggingface.co/roudone/Decider-4B","axes":{"intelligence":27.068221597395038,"calibration":86.57628518330837,"speed":91.8816744041786,"cost":80.74092951874186},"capability":{"score":56.822253390351705,"rank":189,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":155,"api_board_rank":null},"composite_score":16.33572913211934,"rank":228,"open_board_rank":182,"api_board_rank":null,"measured_in":"v1.6.36","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004384994127077292,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 662,682 input + 0 output tokens over 1,485 decisions = 446.250505051 tokens/decision. GPU-time cost USD 0.00438499412707729/1,000; old list-price cost USD 0.0133875151515152/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00438499412707729/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04841596100004608,"p50_s_adjusted":0.24683192200009216,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04841596100004608,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0563412783501235,"p95_s_adjusted":0.262682556700247,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.743016217593315,"capability":56.822253390351705,"axes":{"calibration":86.57628518330837,"cost":66.19900073922123,"intelligence":27.068221597395038,"speed":92.04108849933448},"rank":157,"open_board_rank":119,"api_board_rank":null}},{"key":"rsi-jev-v6.1-vl-27b","name":"RSI-Jev v6.1-VL 27B","board":"open","source_url":"https://huggingface.co/shgao/rsi-jev-v6.1-vl-27b","axes":{"intelligence":61.564156247760415,"calibration":85.34979344070516,"speed":84.6270713604755,"cost":58.679516754522645},"capability":{"score":73.45697484423279,"rank":77,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":55,"api_board_rank":null},"composite_score":70.4002508405647,"rank":92,"open_board_rank":71,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02384237099829211,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 694,381 input + 1,471 output tokens over 1,500 decisions = 463.901333333 tokens/decision. GPU-time cost USD 0.0238423709982921/1,000; old list-price cost USD 0.19736868/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0238423709982921/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13253299049392808,"p50_s_adjusted":0.4150659809878562,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13253299049392808,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.34009330471308197,"p95_s_adjusted":0.830186609426164,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.58850717988564,"capability":73.45697484423279,"axes":{"calibration":85.34979344070516,"cost":31.141652902933714,"intelligence":61.564156247760415,"speed":84.71397705257328},"rank":120,"open_board_rank":87,"api_board_rank":null}},{"key":"ryotide-gemma","name":"RYOTIDE-Gemma (csabag, Gemma-4-E4B stock weights)","board":"open","source_url":"https://github.com/csabag/ryotide","axes":{"intelligence":35.358086407070836,"calibration":73.68644793817732,"speed":89.07386048376334,"cost":null},"capability":{"score":54.52226717262408,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: the frozen pricing path has no market-listing decision for the declared base google/gemma-4-E4B-it and raises rather than estimating; one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 3948e3f1..., and the later re-freezes, which only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base google/gemma-4-E4B-it (the module RAISES 'base model lacks an explicit market-listing decision'); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens as reported by the author's server (input 2,095,834, output_tokens 0)."},"latency":{"p50_s_raw":0.09521084022708237,"p50_s_adjusted":0.34042168045416477,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09521084022708237,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10678864172659815,"p95_s_adjusted":0.3635772834531963,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":55.20857355670594,"axes":{"calibration":74.3509173059117,"cost":null,"intelligence":36.06622980750019,"speed":89.15460718023942},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"ryotide-qwen","name":"RYOTIDE-Qwen (csabag, Qwen3.5-4B stock weights)","board":"open","source_url":"https://github.com/csabag/ryotide","axes":{"intelligence":25.185250272226448,"calibration":87.41983564209619,"speed":89.04172792696406,"cost":76.72111559464565},"capability":{"score":56.30254295716132,"rank":194,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":160,"api_board_rank":null},"composite_score":13.458140852432997,"rank":239,"open_board_rank":192,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005969841675941275,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 902,192 input + 0 output tokens over 1,485 decisions = 607.536700337 tokens/decision. GPU-time cost USD 0.00596984167594128/1,000; old list-price cost USD 0.018226101010101/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00596984167594128/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0967419610824436,"p50_s_adjusted":0.3434839221648872,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0967419610824436,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10650592155288904,"p95_s_adjusted":0.36301184310577805,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":2,"n_over_limit_answered_ok":2,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":13.247205181296051,"capability":56.49884328343206,"axes":{"calibration":87.57576272152932,"cost":62.17918681512502,"intelligence":25.4219238453348,"speed":89.1284523402127},"rank":162,"open_board_rank":123,"api_board_rank":null}},{"key":"ryotide_qwen9","name":"RYOTIDE-Qwen9","board":"open","source_url":"https://github.com/csabag/ryotide/tree/94c71a3d9f044ffaba87981e985b66f47de3be80","axes":{"intelligence":27.627554756423923,"calibration":77.35233849488566,"speed":86.3537180778988,"cost":66.10615580655393},"capability":{"score":52.48994662565479,"rank":210,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":176,"api_board_rank":null},"composite_score":16.10488822146105,"rank":230,"open_board_rank":184,"api_board_rank":null,"measured_in":"v1.6.4","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.01348325681068992,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 901,966 input + 0 output tokens over 1,479 decisions = 609.848546315 tokens/decision. GPU-time cost USD 0.0134832568106899/1,000; old list-price cost USD 0.0487878837052062/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0134832568106899/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13316407600001412,"p50_s_adjusted":0.41632815200002826,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13316407600001412,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.20307530214995495,"p95_s_adjusted":0.5561506042999099,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":21,"n_over_limit_answered_ok":21,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.1347687863636,"capability":53.473263564131784,"axes":{"calibration":78.2699285596189,"cost":49.35064060220574,"intelligence":28.676598568644657,"speed":86.71260028928383},"rank":153,"open_board_rank":115,"api_board_rank":null}},{"key":"sage-1.3.0","name":"Sage 1.3.0 (Levanto Labs)","board":"api","source_url":"https://levanto.ai","axes":{"intelligence":65.59290944525489,"calibration":91.59817606608073,"speed":91.59383325786439,"cost":58.231203787388296},"capability":{"score":78.59554275566781,"rank":46,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":12},"composite_score":73.72797965992304,"rank":49,"open_board_rank":null,"api_board_rank":14,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.024677048070413,"basis":"Levanto list tariff, top tier: USD 0.05 per 1M input tokens, USD 10 per 1M output tokens (https://levanto.ai/pricing, read 5 Oct 2026)"},"latency":{"p50_s_raw":0.17092861235141754,"p50_s_adjusted":0.17092861235141754,"adjustment":"none (API)","p50_s_measured":0.17092861235141754,"latency_basis":"api-end-to-end","p95_s_raw":0.4053236149251461,"p95_s_adjusted":0.4053236149251461,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":74.04992179175468,"capability":78.59554275566781,"axes":{"calibration":91.59817606608073,"cost":58.231203787388296,"intelligence":65.59290944525489,"speed":93.61638181239104},"rank":9,"open_board_rank":null,"api_board_rank":8}},{"key":"sage-1.4.0","name":"Sage 1.4 (Levanto Labs)","board":"api","source_url":"https://levanto.ai/","axes":{"intelligence":63.700734406993774,"calibration":82.91532974968945,"speed":68.89876304791818,"cost":59.47868179714396},"capability":{"score":73.30803207834161,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 2.8× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":67.69827866650344,"rank":118,"open_board_rank":null,"api_board_rank":26,"measured_in":"v1.6.17","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.022423872053872054,"basis":"Levanto Labs public list price, USD 0.05/M input, USD 10/M output (fetched 16 Oct 2026): Levanto's public pricing page lists Input USD 0.05 per 1M tokens on its Free and Developer plans (USD 0.046, 0.042 and 0.038 only on the paid monthly volume plans) and \"All plans: Output $10 / 1M tokens\", where output tokens are Sage's reasoning and answers are free (saved copy sha256 1e98749c...). The page states one tariff for the Sage API and does not name a model version. It is the same list basis the published sage-1.3.0 row carries. Counted as the manufacturer's standard, non-promotional launch list price under interpretation I-1 (the reading that admitted kushal-gemma4-31b-it-autoloops, SubQ Snap and d1-ultra), through a scoped re-freeze of the frozen pricing module that adds exactly three keys (sage-1.4.0, nimble-v4, nimble-v4-max) to its launch-list-price set and nothing else, with a zero-difference proof for every other registry entry and the independent release review; it answers the pricing question that the published review left open. The Free plan's included usage is a trial credit and is not used. ESTIMATE in the scorer's sense: the tariff x this row's OWN provider-reported tokens (665,989 input tokens over the 1,483 answered items; the provider reported 0 output tokens on every answered item, so the output term is zero as measured), divided by the 1,485 items of the common cost basis the scorer uses; the 17 failed requests (15 HTTP 400, 2 HTTP 504) count as zero tokens and as wrong answers. USD 0.022424 per 1,000 decisions on the common basis (USD 0.022200 over all 1,500 items), 0.347x the official cost cap USD 0.064595. Not a bill and not a quote. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.022424 per 1,000 decisions (0.347x the official cost cap USD 0.064595), against USD 0.022200 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":1.705115354496229,"p50_s_adjusted":1.705115354496229,"adjustment":"none (API)","p50_s_measured":1.705115354496229,"latency_basis":"api-end-to-end","p95_s_raw":7.5573568566484015,"p95_s_adjusted":7.5573568566484015,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":13,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.40093468849139,"capability":74.77338506693098,"axes":{"calibration":83.66470689834914,"cost":59.47868179714396,"intelligence":65.88206323551285,"speed":68.80003106294234},"rank":35,"open_board_rank":null,"api_board_rank":19}},{"key":"saif-4b","name":"Saif 4B (Gyr Technology)","board":"open","source_url":"https://huggingface.co/GyrTechnology/saif-4b","axes":{"intelligence":40.11466600558265,"calibration":88.00635147870159,"speed":91.45599093286009,"cost":81.1532909895844},"capability":{"score":64.06050874214212,"rank":137,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":108,"api_board_rank":null},"composite_score":43.23740685453017,"rank":161,"open_board_rank":128,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004248382719727914,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 643,766 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00424838271972791/1,000; old list-price estimate USD 0.012970436534587/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00424838271972791/1,000. Estimated, not a bill. USD 0.00424838272 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00682261406 over all 1,500 items)."},"latency":{"p50_s_raw":0.04333894699999519,"p50_s_adjusted":0.23667789399999037,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.04333894699999519,"latency_basis":"api-end-to-end","p95_s_raw":0.07608217029997207,"p95_s_adjusted":0.3021643405999441,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"scaledown-classify-1","name":"classify-1 (ScaleDown)","board":"api","source_url":"https://scaledown.ai","axes":{"intelligence":21.787775467356628,"calibration":82.82809086203193,"speed":83.72843286582474,"cost":50.78732378203712},"capability":{"score":52.307933164694276,"rank":211,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":34},"composite_score":8.476551950321573,"rank":251,"open_board_rank":null,"api_board_rank":48,"measured_in":"v1.7.3","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.3","measured_on":"2026-10-24","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"mixed","usd_per_1000_decisions":0.04369407407407409,"basis":"ScaleDown vendor list price, USD 0.05/M input tokens; output is free. Explicit vendor reference scaledown:classify-1 (vendor list price), with no base-model price. The frozen pricing module has no key for this API, so this row uses an explicit reference at that price. The scored cost includes the frozen scorer's usage fallback for four HTTP 504 rows: USD 0.04369407407407409 per 1,000 decisions on the common 1,485-item basis (0.67643x the USD 0.06459465517241379 cap); USD 0.082679 per 1,000 over all 1,500. Provider-reported usage alone gives USD 0.035613266 per 1,000 common items and USD 0.074679 over all items, excluding usage estimates for those failures. Cost kind mixed: four rows have estimated usage. Not a bill or quote."},"latency":{"p50_s_raw":0.5046778135074419,"p50_s_adjusted":0.5046778135074419,"adjustment":"none (API)","p50_s_measured":0.5046778135074419,"latency_basis":"api-end-to-end","p95_s_raw":0.8397354750937666,"p95_s_adjusted":0.8397354750937666,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":14,"n_over_limit_answered_ok":10,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"sd-jev-4b-v12.1.0-aligned-s42-step0256","name":"SD-JEV-4B-v12.1.0 (Kego, author-hosted API)","board":"api","source_url":null,"axes":{"intelligence":63.33679285223613,"calibration":86.97988214047655,"speed":81.70557839072836,"cost":83.60967410168716},"capability":{"score":75.15833749635634,"rank":64,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":15},"composite_score":77.69264117305839,"rank":8,"open_board_rank":null,"api_board_rank":4,"measured_in":"v1.8.8","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.8","measured_on":"2026-11-10","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0035183917737814947,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 533,149 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00351839177378149/1,000; old list-price estimate USD 0.0107417528542646/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00351839177378149/1,000. Estimated, not a bill. Customer-stated base Torchcast Decision 4B v1.3 (a Qwen3.5-4B fine-tune) with a LoRA adapter, NOT verified (API-only). Estimate, not a bill and not the author's price; no public tariff exists. Pricing base Qwen/Qwen3.5-4B. The frozen DeepInfra reference was later deprecated; it remains the frozen reference for this estimate. All-items estimate USD 0.00350859401155378/1,000 over all 1,500 items; common-set denominator 1,489. Frozen pricing module sha256 2f52370fe1dba73d75a770a386df350cc76a0898304aeba9e29c91f6e4d3738b."},"latency":{"p50_s_raw":0.5435741060646251,"p50_s_adjusted":0.5435741060646251,"adjustment":"none (API)","p50_s_measured":0.5435741060646251,"latency_basis":"api-end-to-end","p95_s_raw":1.242176762537565,"p95_s_adjusted":1.242176762537565,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.8","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"seb-9b","name":"Seb-9B (ironbcc, Qwen3.5-9B fine-tune, one forward pass, stock vLLM, self-hosted)","board":"open","source_url":"https://huggingface.co/ironbcc/seb-9b","axes":{"intelligence":38.312942577836566,"calibration":87.96306938738428,"speed":90.67341977526846,"cost":68.563136114908},"capability":{"score":63.13800598261042,"rank":147,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":116,"api_board_rank":null},"composite_score":37.2305204893958,"rank":179,"open_board_rank":141,"api_board_rank":null,"measured_in":"v1.6.35","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-20","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011165944447669357,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,494 input + 1,485 output tokens over 1,485 decisions = 505.036363636 tokens/decision. GPU-time cost USD 0.0111659444476694/1,000; old list-price cost USD 0.0505536363636364/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0111659444476694/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.050032105500008583,"p50_s_adjusted":0.2500642110000172,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.050032105500008583,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0962286712500486,"p95_s_adjusted":0.3424573425000972,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":32.56935976147238,"capability":63.13800598261042,"axes":{"calibration":87.96306938738428,"cost":48.88742799396291,"intelligence":38.312942577836566,"speed":90.8001817381692},"rank":95,"open_board_rank":65,"api_board_rank":null}},{"key":"seb-9b-v2","name":"Seb-9B-v2 (ironbcc, Qwen3.5-9B fine-tune, one forward pass, stock vLLM, self-hosted)","board":"open","source_url":"https://huggingface.co/ironbcc/Seb-9B-v2","axes":{"intelligence":46.11278573865326,"calibration":83.5320529378626,"speed":91.9752040844015,"cost":68.563136114908},"capability":{"score":64.82241933825793,"rank":134,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":106,"api_board_rank":null},"composite_score":57.552603401286994,"rank":134,"open_board_rank":104,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011165944447669357,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 9B active parameters; throughput 27344.4444444 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 748,494 input + 1,485 output tokens over 1,485 decisions = 505.036363636 tokens/decision. GPU-time cost USD 0.0111659444476694/1,000; old list-price cost USD 0.0505536363636364/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0111659444476694/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03548585999999432,"p50_s_adjusted":0.22097171999998863,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03548585999999432,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.06858626535004646,"p95_s_adjusted":0.2871725307000929,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"semif-qwen3.5-4b","name":"SemIf, formerly OpenJev (Qwen3.5-4B, TheoLeeCJ)","board":"open","source_url":"https://github.com/TheoLeeCJ/openjev","axes":{"intelligence":27.10113669038628,"calibration":83.59960184071504,"speed":90.43707331448911,"cost":77.61552273233744},"capability":{"score":55.35036926555066,"rank":199,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":165,"api_board_rank":null},"composite_score":16.141782263417074,"rank":229,"open_board_rank":183,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005573772548130538,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 921,181 input + 0 output tokens over 1,624 decisions = 567.229679803 tokens/decision. GPU-time cost USD 0.00557377254813054/1,000; old list-price cost USD 0.0170168903940887/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00557377254813054/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05878807799308561,"p50_s_adjusted":0.26757615598617124,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05878807799308561,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0939722015406005,"p95_s_adjusted":0.337944403081201,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":15.517179498801362,"capability":55.35036926555066,"axes":{"calibration":83.59960184071504,"cost":63.07359395281681,"intelligence":27.10113669038628,"speed":90.76560119421038},"rank":159,"open_board_rank":121,"api_board_rank":null}},{"key":"shisa-de-1","name":"shisa-de-1 (shisa-ai, Gemma-4-26B-A4B)","board":"open","source_url":"https://huggingface.co/shisa-ai/shisa-de-1","axes":{"intelligence":49.90370340622748,"calibration":81.16385348785812,"speed":92.69330974783162,"cost":71.1478434006942},"capability":{"score":65.5337784470428,"rank":130,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":102,"api_board_rank":null},"composite_score":69.6574933188226,"rank":102,"open_board_rank":80,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009156690883981826,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,620 input + 0 output tokens over 1,485 decisions = 524.996632997 tokens/decision. GPU-time cost USD 0.00915669088398183/1,000; old list-price cost USD 0.047249696969697/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00915669088398183/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03166249021887779,"p50_s_adjusted":0.21332498043775558,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03166249021887779,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.051065808977000415,"p95_s_adjusted":0.25213161795400085,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":62.425924800405696,"capability":65.54874380543649,"axes":{"calibration":81.19978921578861,"cost":49.76802917325137,"intelligence":49.89769839508435,"speed":92.81832911721798},"rank":54,"open_board_rank":32,"api_board_rank":null}},{"key":"sieve-31b-v2","name":"Sieve-31B-v2","board":"open","source_url":"https://huggingface.co/sthanika-ai/Sieve-31B-v2","axes":{"intelligence":68.69375926172683,"calibration":93.43694381196171,"speed":85.61790755169669,"cost":56.225182205710844},"capability":{"score":81.06535153684428,"rank":30,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":22,"api_board_rank":null},"composite_score":73.0924793557567,"rank":60,"open_board_rank":45,"api_board_rank":null,"measured_in":"v1.7.5","draw":{"id":"v1.6-regular-20261012","label":null,"release":"v1.7.5","measured_on":"2026-10-25","status":null},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.028784626140763356,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 714,573 input + 0 output tokens over 1,485 decisions = 481.193939394 tokens/decision. GPU-time cost USD 0.0287846261407634/1,000; old list-price cost USD 0.0433074545454545/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0287846261407634/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.13763563850000082,"p50_s_adjusted":0.42527127700000167,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.13763563850000082,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.24748773194991144,"p95_s_adjusted":0.6449754638998229,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.5","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"sieve-31b-v3","name":"Sieve-31B-v3","board":"open","source_url":"https://huggingface.co/sthanika-ai/Sieve-31B-v3","axes":{"intelligence":77.42850447479417,"calibration":92.43750988443877,"speed":83.06216589785838,"cost":52.454582600169914},"capability":{"score":84.93300717961647,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":72.94405398679245,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.16","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.16","measured_on":"2026-11-19","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.038445653430234025,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.038445653430234/1,000; old list-price estimate USD 0.0578428005372733/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.038445653430234/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.18854752100003225,"p50_s_adjusted":0.5270950420000645,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.18854752100003225,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.3936671171498574,"p95_s_adjusted":0.9373342342997149,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.16","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"simplejev-qwen3.5-0.8b","name":"SimpleJev (Qwen3.5-0.8B, CPU)","board":"open","source_url":"https://github.com/featherless-ai/simple-jev","axes":{"intelligence":12.67752479561282,"calibration":49.46821562395294,"speed":58.64501925086723,"cost":76.868497665666},"capability":{"score":31.07287020978288,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 14.9× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":1.991018277559104,"rank":269,"open_board_rank":218,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005902691392505279,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,585,602 input + 0 output tokens over 1,624 decisions = 976.35591133 tokens/decision. GPU-time cost USD 0.00590269139250528/1,000; old list-price cost USD 0.00976355911330049/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00590269139250528/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":4.505720648681745,"p50_s_adjusted":9.16144129736349,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":4.505720648681745,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":7.380972131667658,"p95_s_adjusted":14.911944263335316,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":14,"n_over_limit_answered_ok":14,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":2.112084174105171,"capability":31.59915405528791,"axes":{"calibration":50.197881904888476,"cost":70.31175519898045,"intelligence":13.000426205687345,"speed":58.929227255635794},"rank":197,"open_board_rank":155,"api_board_rank":null}},{"key":"simplejev-qwen3.6-35b-a3b","name":"SimpleJev Qwen3.6-35B-A3B","board":"api","source_url":"https://github.com/featherless-ai/simple-jev","axes":{"intelligence":56.29,"calibration":80.3,"speed":76.09,"cost":35.15},"capability":{"score":68.295,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 4.5× Jev (v1.5 reference)","latency 2.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":27.53,"rank":197,"open_board_rank":null,"api_board_rank":41,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.1451164454277286,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: exact base-model market reference (operator tariff excluded by 30-day rule)"},"latency":{"p50_s_raw":0.652,"p50_s_adjusted":1.304,"adjustment":"x2 (demo)","p50_s_measured":0.652,"latency_basis":"demo-endpoint-x2","p95_s_raw":0.943,"p95_s_adjusted":1.885,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":27.57,"capability":68.29,"axes":{"calibration":80.3,"cost":35.15,"intelligence":56.29,"speed":76.7},"rank":106,"open_board_rank":null,"api_board_rank":33}},{"key":"simplejev-qwen3.8-27b","name":"SimpleJev Qwen3.8-27B","board":"api","source_url":"https://github.com/featherless-ai/simple-jev","axes":{"intelligence":52.88,"calibration":90.33,"speed":76.44,"cost":14.9},"capability":{"score":71.605,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 21.3× Jev (v1.5 reference)","latency 2.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":3.22,"rank":266,"open_board_rank":null,"api_board_rank":50,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.686803448275862,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: exact base-model market reference (operator tariff excluded by 30-day rule)"},"latency":{"p50_s_raw":0.633,"p50_s_adjusted":1.266,"adjustment":"x2 (demo)","p50_s_measured":0.633,"latency_basis":"demo-endpoint-x2","p95_s_raw":0.897,"p95_s_adjusted":1.794,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.22,"capability":71.6,"axes":{"calibration":90.33,"cost":14.9,"intelligence":52.88,"speed":76.93},"rank":193,"open_board_rank":null,"api_board_rank":41}},{"key":"simplejev-qwen3.8-27b-selfhost","name":"SimpleJev Qwen3.8-27B (self-hosted, v1.6 pool)","board":"open","source_url":"https://github.com/featherless-ai/simple-jev","axes":{"intelligence":57.92931033189586,"calibration":91.85610710600702,"speed":75.96701882114468,"cost":42.47169570460766},"capability":{"score":74.89270871895144,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.6× Jev (v1.5 reference)","latency 2.3× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":44.49974984123803,"rank":153,"open_board_rank":121,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.08271992713196244,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,377,206 input + 0 output tokens over 1,477 decisions = 1609.48273527 tokens/decision. GPU-time cost USD 0.0827199271319624/1,000; old list-price cost USD 0.675982748815166/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0827199271319624/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.63900578051107,"p50_s_adjusted":1.42801156102214,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.63900578051107,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.8112094948810408,"p95_s_adjusted":1.7724189897620815,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.394848880196433,"capability":74.89270871895144,"axes":{"calibration":91.85610710600702,"cost":15.10193161538902,"intelligence":57.92931033189586,"speed":76.19964460590472},"rank":189,"open_board_rank":149,"api_board_rank":null}},{"key":"singularity-super-decider","name":"Super Decider","board":"api","source_url":"https://singularitycompute.com/super-decider","axes":{"intelligence":78.8866847061141,"calibration":90.84828412834014,"speed":85.30638666270315,"cost":64.26424619973979},"capability":{"score":84.86748441722712,"rank":9,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":3},"composite_score":78.48219079083019,"rank":4,"open_board_rank":null,"api_board_rank":2,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"listed","usd_per_1000_decisions":0.01553072727272727,"basis":"Launch list price: the launch list price stated on the vendor's public product page (singularitycompute.com/super-decider, fetched 22 Oct 2026: USD 0.03 per 1M input tokens, output tokens free); the ASI:Cloud catalogue itself was not reachable to us. The page lists one price for the product and no separate rate for thinking mode; we apply it to both rows. Counted as an operator launch list price by the frozen pricing module pricing_v15 (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen on 22 Oct 2026 with launch-list keys added and nothing else changed; the launch-list line and the rates for this product are identical in the earlier module 81e3578f), which carries a launch-list-price line for this product. Applied to the tokens reported by the provider's API; not a base-model estimate. Over all 1,500 items the frozen resolve() gives USD 0.039197 per 1,000 decisions; the official cost axis uses the common 1,485-item basis, where it is USD 0.015531 per 1,000 decisions. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 768,771 input / 322,546 output tokens. All-item usage 1,959,826 input / 328,348 output over 1,500 items."},"latency":{"p50_s_raw":0.24612953246105462,"p50_s_adjusted":0.24612953246105462,"adjustment":"none (API)","p50_s_measured":0.24612953246105462,"latency_basis":"api-end-to-end","p95_s_raw":1.1972851804515812,"p95_s_adjusted":1.1972851804515812,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":11,"n_over_limit_answered_ok":11,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"singularity-super-decider-thinking","name":"Super Decider (thinking)","board":"api","source_url":"https://singularitycompute.com/super-decider","axes":{"intelligence":80.00899811568502,"calibration":91.69553288708067,"speed":79.83008200555805,"cost":56.397623391206686},"capability":{"score":85.85226550138285,"rank":5,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":2},"composite_score":74.54050341581339,"rank":35,"open_board_rank":null,"api_board_rank":12,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-21","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"listed","usd_per_1000_decisions":0.028406161616161614,"basis":"Launch list price: the launch list price stated on the vendor's public product page (singularitycompute.com/super-decider, fetched 22 Oct 2026: USD 0.03 per 1M input tokens, output tokens free); the ASI:Cloud catalogue itself was not reachable to us. The page lists one price for the product and no separate rate for thinking mode; we apply it to both rows. Counted as an operator launch list price by the frozen pricing module pricing_v15 (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen on 22 Oct 2026 with launch-list keys added and nothing else changed; the launch-list line and the rates for this product are identical in the earlier module 81e3578f), which carries a launch-list-price line for this product. Applied to the tokens reported by the provider's API; not a base-model estimate. Over all 1,500 items the frozen resolve() gives USD 0.052098 per 1,000 decisions; the official cost axis uses the common 1,485-item basis, where it is USD 0.028406 per 1,000 decisions. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 1,406,105 input / 636,221 output tokens. All-item usage 2,604,915 input / 643,964 output over 1,500 items."},"latency":{"p50_s_raw":0.8880362859927118,"p50_s_adjusted":0.8880362859927118,"adjustment":"none (API)","p50_s_measured":0.8880362859927118,"latency_basis":"api-end-to-end","p95_s_raw":1.171011304546846,"p95_s_adjusted":1.171011304546846,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":4,"n_over_limit_answered_ok":4,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"smalljev","name":"smalljev semantic-v9","board":"open","source_url":"https://github.com/isHeSatoshi/smalljev","axes":{"intelligence":8.670533657732655,"calibration":77.23092648324078,"speed":90.11972690216095,"cost":60.77469204714053},"capability":{"score":42.95073007048671,"rank":243,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":206,"api_board_rank":null},"composite_score":0.7718897362808083,"rank":273,"open_board_rank":221,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020300665024630543,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.06696509593166411,"p50_s_adjusted":0.28393019186332824,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06696509593166411,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09631119249388576,"p95_s_adjusted":0.34262238498777153,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.7719455446134506,"capability":42.95073007048671,"axes":{"calibration":77.23092648324078,"cost":60.77469204714053,"intelligence":8.670533657732655,"speed":90.21131723708834},"rank":200,"open_board_rank":157,"api_board_rank":null}},{"key":"sors-2b","name":"Sors-2B (SakuraYuyuko)","board":"open","source_url":"https://huggingface.co/SakuraYuyuko/Sors-2B","axes":{"intelligence":31.350346976303964,"calibration":76.5498355413793,"speed":93.09919217560679,"cost":91.07236654889317},"capability":{"score":53.95009125884163,"rank":208,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":174,"api_board_rank":null},"composite_score":23.58267194009419,"rank":206,"open_board_rank":165,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-30","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.001984210674576524,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 488,698 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00198421067457652/1,000; old list-price estimate USD 0.00656411014103425/1,000 (deepinfra:Qwen/Qwen3.5-2B at USD 0.02/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00198421067457652/1,000. Estimated, not a bill. USD 0.00198421067 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00197894186 over all 1,500 items)."},"latency":{"p50_s_raw":0.03263867649999952,"p50_s_adjusted":0.21527735299999903,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.03263867649999952,"latency_basis":"api-end-to-end","p95_s_raw":0.03877646650001907,"p95_s_adjusted":0.22755293300003815,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"spark-s1-4b-v6","name":"spark-s1-4b-v6 (Open Spark Jev, abhishek085)","board":"open","source_url":"https://github.com/abhishek085/open-spark-jev","axes":{"intelligence":43.2364603010071,"calibration":64.79562625454915,"speed":86.33664750658791,"cost":74.96399197637558},"capability":{"score":54.016043277778124,"rank":206,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":172,"api_board_rank":null},"composite_score":47.114595108116106,"rank":146,"open_board_rank":114,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006831775724763977,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,129,092 input + 0 output tokens over 1,624 decisions = 695.253694581 tokens/decision. GPU-time cost USD 0.00683177572476398/1,000; old list-price cost USD 0.0208576108374384/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00683177572476398/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12928083550286829,"p50_s_adjusted":0.4085616710057366,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12928083550286829,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.20947730995176245,"p95_s_adjusted":0.5689546199035249,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":44.90217636528639,"capability":54.016043277778124,"axes":{"calibration":64.79562625454915,"cost":60.422063196854964,"intelligence":43.2364603010071,"speed":86.95618624402076},"rank":71,"open_board_rank":46,"api_board_rank":null}},{"key":"spx-cd-flash","name":"SPX-CD Flash (SurdAI, hosted /v1/systemone, Oct-4 checkpoint)","board":"api","source_url":"https://surd.ai","axes":{"intelligence":58.43051843707937,"calibration":90.20036269195566,"speed":79.18845201110972,"cost":52.107175358413464},"capability":{"score":74.31544056451752,"rank":68,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":18},"composite_score":66.64434875110587,"rank":122,"open_board_rank":null,"api_board_rank":27,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.039484576844955994,"basis":"Operator's own public-beta overage tariff, conservative reading USD 0.04 per 1M input tokens, output USD 0 (the provider's /v1/models payload reports output_price \"0\"): three dated primary readings of the operator's own published overage tariff: 24 Sep 2026 provider console (flash 0.025, pro 0.08 per 1M input), 1 Oct 2026 docs page (flash 0.04, pro 0.02), 6 Oct 2026 docs page AND /v1/models payload agreeing (flash 0.02, pro 0.04). Sources conflict, so the per-model MAXIMUM is used, which is the published fastino-gliner-2-5-decide precedent (\"published sources conflict ... conservative higher\"). Today's agreeing pair would be USD 0.02/M, i.e. half this row's cost. No base-model floor applies: the submitter states the base model stays sealed until a later open release, so no base model exists to reference. Priced on usage.input_tokens as the provider reports them. The provider also reports billable_input_tokens = input_tokens / 2 on every single item (297/297 of the public set, exactly 2.0): its default effort 2 renders the prompt twice and its docs bundle the second pass free during public beta (\"default effort 2 is included at the single-pass input price\"). That bundle is a public-beta promotion, and TASK.md forbids a free or promotional tier setting the price, so the two passes actually consumed are charged. The operator's own billed figure is exactly half of this row's cost"},"latency":{"p50_s_raw":1.0145484656095505,"p50_s_adjusted":1.0145484656095505,"adjustment":"none (API)","p50_s_measured":1.0145484656095505,"latency_basis":"api-end-to-end","p95_s_raw":1.18817935064435,"p95_s_adjusted":1.18817935064435,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":66.68017155386212,"capability":74.31544056451752,"axes":{"calibration":90.20036269195566,"cost":52.107175358413464,"intelligence":58.43051843707937,"speed":79.39117060033084},"rank":44,"open_board_rank":null,"api_board_rank":21}},{"key":"spx-cd-omni","name":"SPX-CD-Omni (SurdAI, google/gemma-4-12B-it + LoRA r32, one forward pass, self-hosted)","board":"open","source_url":"https://huggingface.co/SurdAI/SPX-CD-Omni","axes":{"intelligence":48.779308810848946,"calibration":87.47801599949784,"speed":89.06649387039909,"cost":68.67068018166556},"capability":{"score":68.1286624051734,"rank":120,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":93,"api_board_rank":null},"composite_score":65.95560252970363,"rank":125,"open_board_rank":98,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011074156600503763,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 726,752 input + 0 output tokens over 1,477 decisions = 492.046039269 tokens/decision. GPU-time cost USD 0.0110741566005038/1,000; old list-price cost USD 0.0246023019634394/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0110741566005038/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.07505282340571284,"p50_s_adjusted":0.3001056468114257,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.07505282340571284,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.13156010708771645,"p95_s_adjusted":0.4131202141754329,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":63.16406085537614,"capability":68.1286624051734,"axes":{"calibration":87.47801599949784,"cost":58.27072766098771,"intelligence":48.779308810848946,"speed":89.44891259349035},"rank":51,"open_board_rank":30,"api_board_rank":null}},{"key":"spx-cd-pro","name":"SPX-CD Pro (SurdAI, hosted /v1/systemone, Oct-4 checkpoint)","board":"api","source_url":"https://surd.ai","axes":{"intelligence":64.6277726176258,"calibration":92.33141939321372,"speed":77.43431239288299,"cost":43.076275488494026},"capability":{"score":78.47959600541975,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.4× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":47.553774484141584,"rank":144,"open_board_rank":null,"api_board_rank":32,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.07896915368991199,"basis":"Operator's own public-beta overage tariff, conservative reading USD 0.08 per 1M input tokens, output USD 0: three dated primary readings of the operator's own published overage tariff: 24 Sep 2026 provider console (flash 0.025, pro 0.08 per 1M input), 1 Oct 2026 docs page (flash 0.04, pro 0.02), 6 Oct 2026 docs page AND /v1/models payload agreeing (flash 0.02, pro 0.04). Sources conflict, so the per-model MAXIMUM is used, which is the published fastino-gliner-2-5-decide precedent (\"published sources conflict ... conservative higher\"). Today's agreeing pair would be USD 0.04/M, i.e. half this row's cost. No base-model floor applies: the submitter states the base model stays sealed until a later open release. Priced on usage.input_tokens as the provider reports them. The provider also reports billable_input_tokens = input_tokens / 2 on every single item (297/297 of the public set, exactly 2.0): its default effort 2 renders the prompt twice and its docs bundle the second pass free during public beta (\"default effort 2 is included at the single-pass input price\"). That bundle is a public-beta promotion, and TASK.md forbids a free or promotional tier setting the price, so the two passes actually consumed are charged. The operator's own billed figure is exactly half of this row's cost"},"latency":{"p50_s_raw":1.1453867703676224,"p50_s_adjusted":1.1453867703676224,"adjustment":"none (API)","p50_s_measured":1.1453867703676224,"latency_basis":"api-end-to-end","p95_s_raw":1.576219144836068,"p95_s_adjusted":1.576219144836068,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":47.66970820706367,"capability":78.63514339971708,"axes":{"calibration":92.41334857032513,"cost":43.076275488494026,"intelligence":64.85693822910903,"speed":77.96292698098645},"rank":68,"open_board_rank":null,"api_board_rank":25}},{"key":"standard_one_3b_sh","name":"Standard One 3B SH (Standard Thinking)","board":"open","source_url":"https://huggingface.co/StandardThinking/StandardOne-3B-SH","axes":{"intelligence":32.78619016656576,"calibration":78.35523487828182,"speed":86.38605449679788,"cost":48.64218899335634},"capability":{"score":55.57071252242379,"rank":197,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":163,"api_board_rank":null},"composite_score":21.588999608530766,"rank":209,"open_board_rank":168,"api_board_rank":null,"measured_in":"v1.6.10","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.05151420875420875,"basis":"v1.5-M2 base-model reference (read 2026-10-12T00:14Z, source/public-docs/OPENROUTER-MINISTRAL-3-PRICES.json): the model card names base_model StandardThinking/StandardOne-3B, built on mistralai/Ministral-3-3B-Instruct-2512. Neither Standard One checkpoint is listed on OpenRouter; the exact underlying base is: OpenRouter mistralai/ministral-3b-2512 list price USD 0.10/M input and USD 0.10/M output. Self-hosted weights, no bookable tariff of the measured deployment; the author's stated hosted list price (USD 0.038/M input, standardthinking.ai/pricing) is below this floor, so the M2 §2 floor binds (price floor: base-model reference price applied). The schema head generates no text, so output tokens are 0 and priced at the reference output price. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.051514 per 1,000 decisions (0.797x the official cost cap USD 0.064595), against USD 0.050999 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.03863196854945272,"p50_s_adjusted":0.22726393709890544,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03863196854945272,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.43063139129662886,"p95_s_adjusted":1.0112627825932576,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.595754998730687,"capability":55.57071252242379,"axes":{"calibration":78.35523487828182,"cost":48.64218899335634,"intelligence":32.78619016656576,"speed":86.56241878672063},"rank":119,"open_board_rank":86,"api_board_rank":null}},{"key":"standard_one_3b_sh_v4","name":"Standard One 3B SH v4 (Standard Thinking)","board":"open","source_url":"https://huggingface.co/StandardThinking/StandardOne-3B-SH","axes":{"intelligence":34.357350203953175,"calibration":77.72284683250058,"speed":93.10096025009605,"cost":48.64218899335634},"capability":{"score":56.04009851822688,"rank":196,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":162,"api_board_rank":null},"composite_score":24.395574588088643,"rank":205,"open_board_rank":164,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.05151420875420875,"basis":"ESTIMATE (*est.), base-model market reference, the same estimate as the live row standard_one_3b_sh: the model card names base_model StandardThinking/StandardOne-3B, built on mistralai/Ministral-3-3B-Instruct-2512. Neither Standard One checkpoint is listed on OpenRouter; the exact underlying base is priced as OpenRouter mistralai/ministral-3b-2512, list price USD 0.10/M input and USD 0.10/M output (the same OpenRouter document the live sibling cites, read 2026-10-12T00:14Z, source/public-docs/OPENROUTER-MINISTRAL-3-PRICES.json; the frozen module catalogue carries the same 0.10/0.10). Self-hosted weights, no bookable tariff of the measured deployment; the author's stated hosted list price (USD 0.038/M input, standardthinking.ai/pricing) is below this floor, so the base-model reference price applies as a floor. Note on the module: the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen 22 Oct 2026) has no explicit listing decision for this base string (reference() raises), so the price is the registry value 0.10/0.10 carried over from the live entry and verified against that snapshot, not a module lookup. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the usage block of the author's own server: input_tokens = 764,986 on the 1,485 answered items, output_tokens = 0 on every row (the schema head generates no text; priced at the reference output price). Our offline count through a first-party transcription of the author's renderer and the pinned tokenizer predicted the server's numbers EXACTLY, item by item and in total (764,986 offline = 764,986 native), and equals the 12 Oct receipt of the v2 row (same tokenizer, template and item set). The 15 refused items (HTTP 422, over the author's 32,768-token context) count 0 tokens."},"latency":{"p50_s_raw":0.030541504000012765,"p50_s_adjusted":0.21108300800002552,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.030541504000012765,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.040990041799960866,"p95_s_adjusted":0.23198008359992173,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"standard_one_8b_sh","name":"Standard One 8B SH (Standard Thinking)","board":"open","source_url":"https://huggingface.co/StandardThinking/StandardOne-8B-SH","axes":{"intelligence":34.464905061466695,"calibration":85.49831885164681,"speed":84.17488059602204,"cost":43.21124696592469},"capability":{"score":59.981611956556755,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.4× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":18.74307320095725,"rank":217,"open_board_rank":174,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0781553,"basis":"v1.5-M2 base-model reference: model card base_model StandardThinking/StandardOne-8B, built on mistralai/Ministral-3-8B-Instruct-2512 (not listed on OpenRouter as Standard One). OpenRouter mistralai/ministral-8b-2512 list price USD 0.15/M input and USD 0.15/M output, read 2026-10-10T08:43Z from https://openrouter.ai/api/v1/models. Schema head generates no text: 0 output tokens."},"latency":{"p50_s_raw":0.06744719750713557,"p50_s_adjusted":0.28489439501427116,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06744719750713557,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.5961166904016864,"p95_s_adjusted":1.3422333808033726,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":18.74677558938401,"capability":59.981611956556755,"axes":{"calibration":85.49831885164681,"cost":43.21124696592469,"intelligence":34.464905061466695,"speed":84.28099014807135},"rank":138,"open_board_rank":103,"api_board_rank":null}},{"key":"standardone-8b","name":"Standard One 8B (Standard Thinking)","board":"open","source_url":"https://huggingface.co/StandardThinking/StandardOne-8B","axes":{"intelligence":32.80426604507879,"calibration":84.01508092906352,"speed":92.46998352781172,"cost":43.283250331904235},"capability":{"score":58.40967348707115,"rank":null,"eligible":false,"within_caps":false,"reasons":["cost 2.4× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":16.909498058775554,"rank":225,"open_board_rank":180,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.07772456896551724,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate"},"latency":{"p50_s_raw":0.03300525527447462,"p50_s_adjusted":0.21601051054894924,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03300525527447462,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.056068028882145876,"p95_s_adjusted":0.2621360577642917,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.917108299009165,"capability":58.40967348707115,"axes":{"calibration":84.01508092906352,"cost":43.283250331904235,"intelligence":32.80426604507879,"speed":92.76442849252426},"rank":149,"open_board_rank":112,"api_board_rank":null}},{"key":"starlord-27b-v2-0-preview","name":"Starlord-27B-v2.0-preview","board":"open","source_url":"https://huggingface.co/xsdfasfgsa/Starlord-27B-v2.0-preview","axes":{"intelligence":75.4315147322705,"calibration":93.52883876482899,"speed":88.13780157140738,"cost":61.067776822568604},"capability":{"score":84.48017674854975,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":77.414262116208,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.12","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.12","measured_on":"2026-11-15","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.01984909784559396,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.019849097845594/1,000; old list-price estimate USD 0.162205748824715/1,000 (qwen/qwen3.8-27b at USD 0.42/M input and USD 3.0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.019849097845594/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.08506945850001557,"p50_s_adjusted":0.32013891700003116,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08506945850001557,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.1648012131000371,"p95_s_adjusted":0.4796024262000742,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.12","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"starlord-4b-preview","name":"Starlord-4B-preview","board":"open","source_url":"https://huggingface.co/xsdfasfgsa/Starlord-4B-preview","axes":{"intelligence":58.647519213382566,"calibration":71.24809230925125,"speed":92.11986108726356,"cost":73.76266219513131},"capability":{"score":64.94780576131691,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":72.07350067955714,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.15","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.15","measured_on":"2026-11-17","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Added after the v1.8.10 ranking review; it is ranked at the next review.","price":{"kind":"estimate","usd_per_1000_decisions":0.0074916577162910765,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.00749165771629108/1,000; old list-price estimate USD 0.0228722498321021/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00749165771629108/1,000. Estimated, not a bill."},"latency":{"p50_s_raw":0.03078817400000844,"p50_s_adjusted":0.21157634800001687,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03078817400000844,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0700496813499626,"p95_s_adjusted":0.2900993626999252,"n":null},"method":"v1.7","rescored_in":null,"new_in":"v1.8.15","method_status":"v1.7","decision_limit":{"limit_s":10,"rule":"latency_s > limit -> failed","treatment":"share"},"previous":null},{"key":"startlux-decision-v1.1-27b-autothink","name":"StartLux-Decision v1.1 27B AutoThink (StartLux Labs, Qwen3.8-27B fine-tune, gate 0.6)","board":"open","source_url":"https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-AutoThink","axes":{"intelligence":66.69999992075427,"calibration":90.90819924197592,"speed":75.15747300434305,"cost":56.56867594442542},"capability":{"score":78.8040995813651,"rank":45,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":34,"api_board_rank":null},"composite_score":70.20544677571385,"rank":94,"open_board_rank":73,"api_board_rank":null,"measured_in":"v1.6.25","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02803566112177389,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 737,990 input + 72,063 output tokens over 1,485 decisions = 545.49023569 tokens/decision. GPU-time cost USD 0.0280356611217739/1,000; old list-price cost USD 0.354306262626263/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0280356611217739/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06987204699544236,"p50_s_adjusted":0.28974409399088474,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06987204699544236,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":5.187693915035915,"p95_s_adjusted":10.52538783007183,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":12,"n_over_limit_answered_ok":12,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":10.892954017975724,"capability":80.34729259324169,"axes":{"calibration":91.73088868169317,"cost":23.51863513364647,"intelligence":68.96369650479022,"speed":75.05228443363347},"rank":168,"open_board_rank":129,"api_board_rank":null}},{"key":"startlux-decision-v1.1-27b-instant","name":"StartLux-Decision v1.1 27B Instant (StartLux Labs, Qwen3.8-27B fine-tune)","board":"open","source_url":"https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-Instant","axes":{"intelligence":73.19432958384127,"calibration":91.92111224732803,"speed":90.32360333729886,"cost":57.87760156376582},"capability":{"score":82.55772091558464,"rank":16,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":11,"api_board_rank":null},"composite_score":75.62823658595717,"rank":24,"open_board_rank":15,"api_board_rank":null,"measured_in":"v1.6.25","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.025355953571322483,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11849.1171499 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=base-arch). The row's own scored cost basis: 737,990 input + 0 output tokens over 1,485 decisions = 496.962962963 tokens/decision. GPU-time cost USD 0.0253559535713225/1,000; old list-price cost USD 0.208724444444444/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0253559535713225/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.05708817200502381,"p50_s_adjusted":0.26417634401004764,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.05708817200502381,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10067733919713646,"p95_s_adjusted":0.3513546783942729,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.618048863688184,"capability":82.55772091558464,"axes":{"calibration":91.92111224732803,"cost":30.41280058663709,"intelligence":73.19432958384127,"speed":90.62312905261143},"rank":118,"open_board_rank":85,"api_board_rank":null}},{"key":"strike-1","name":"Strike-1","board":"api","source_url":"https://strike-ai.eu","axes":{"intelligence":42.46302428520289,"calibration":89.50425071284391,"speed":83.29564724453803,"cost":54.05958090666912},"capability":{"score":65.9836374990234,"rank":129,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":27},"composite_score":44.230359262377306,"rank":157,"open_board_rank":null,"api_board_rank":33,"measured_in":"v1.7.16","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.16","measured_on":"2026-10-28","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.03398969778374748,"basis":"ESTIMATE (*est.), hosted API without a counted list price: the vendor's stated base Qwen/Qwen3.5-4B at its market reference deepinfra:Qwen/Qwen3.5-4B, USD 0.03 per million input tokens and USD 0.15 per million output tokens (frozen pricing module sha256 645fd82eac728d597048bec0995ecd3ac896566e269fb6c9a1fdb84c171a9ac4), x the vendor-reported usage: 1,687,022 input tokens and 0 output tokens on the 1,489 common items (output 0 as reported: no text is generated). COST FIGURES: USD 0.033990 per 1,000 decisions, 0.53x the official cost cap USD 0.064595. Cost on the 1,489 items every row on this draw answered (same-items rule); all-items cost USD 0.033853 per 1,000 decisions. Estimated model tariff, not charged usage or a host quote; the vendor reported no charge for these requests."},"latency":{"p50_s_raw":0.47458370300591923,"p50_s_adjusted":0.47458370300591923,"adjustment":"none (API)","p50_s_measured":0.47458370300591923,"latency_basis":"api-end-to-end","p95_s_raw":0.9865576189680723,"p95_s_adjusted":0.9865576189680723,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.16","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"subq-snap-1.1","name":"SubQ Snap 1.1 (SubQ)","board":"api","source_url":"https://subq.ai","axes":{"intelligence":75.75710514449688,"calibration":88.06513871774217,"speed":84.90104208468355,"cost":60.67742028447466},"capability":{"score":81.91112193111952,"rank":20,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":6},"composite_score":75.73702535215112,"rank":21,"open_board_rank":null,"api_board_rank":8,"measured_in":"v1.6.43","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020452794612794613,"basis":"Vendor's published standard launch tariff, USD 0.04 per million input tokens; output free (read 22 Oct 2026 at https://docs.subq.ai/using-the-api/models/ and https://subq.ai). Counted under the frozen I-1 launch-list-price rule. Capacity-limited public preview, first come first served, with a waitlist once capacity is reached; not openly bookable by everyone. Promotional signup credits are excluded. Own measured usage: 1,939,205 input / 0 output tokens over all 1,500 decisions. Common 1,485-item cost basis: 759,310 input / 0 output tokens, USD 0.020452795 per 1,000 decisions; all-item cost USD 0.051712133 per 1,000. The historical preview row keeps its frozen USD 0.05/M input tariff."},"latency":{"p50_s_raw":0.4690868764882907,"p50_s_adjusted":0.4690868764882907,"adjustment":"none (API)","p50_s_measured":0.4690868764882907,"latency_basis":"api-end-to-end","p95_s_raw":0.6896718636329751,"p95_s_adjusted":0.6896718636329751,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":75.7463117322681,"capability":81.91112193111952,"axes":{"calibration":88.06513871774217,"cost":60.67742028447466,"intelligence":75.75710514449688,"speed":84.94774043787051},"rank":5,"open_board_rank":null,"api_board_rank":4}},{"key":"subq-snap-preview","name":"SubQ Snap (preview, SubQ)","board":"api","source_url":"https://subq.ai","axes":{"intelligence":72.25005247787018,"calibration":90.69740052356492,"speed":86.35826133612204,"cost":57.693900899055016},"capability":{"score":81.47372650071756,"rank":28,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":8},"composite_score":74.37845073677168,"rank":39,"open_board_rank":null,"api_board_rank":13,"measured_in":"v1.6.17","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-13","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.025715993265993265,"basis":"Operator's own published standard usage tariff: USD 0.05 per 1M INPUT tokens with no output billing. Two vendor sources, both snapshotted: the pricing section of https://subq.ai/ read on 14 Oct 2026 08:10Z ('In $0.05 per 1M tokens Input / Out Free Output') and the vendor's own launch post 'Introducing SubQ Snap' ('SubQ Snap is priced at $0.05 per million input tokens, with no charge for output tokens'), which also dates the tariff to the product's launch; Subquadratic's Alexander Whedon separately confirmed the figure in writing to the evaluator on 14 Oct 2026 01:16 Berlin (the published review). Counted under interpretation I-1 (a manufacturer's standard, non-promotional launch list price counts from day 1 - the reading that already admitted kushal-gemma4-31b-it-autoloops). DISCLOSED: Snap is in a capacity-limited public preview - first come, first served, with a waitlist once capacity is reached - so this is the vendor's stated usage price for a product that is not yet openly bookable by everyone; the published pricing interpretation states explicitly that I-1 covers this tariff during that public preview, on the condition that the capacity limit and waitlist are stated here, as they are. The same pricing section offers 'Free 100M tokens for the first 10,000 users': that is a promotional trial credit and is deliberately NOT used - a free tier is priced at the published paid tariff, never as cost 0. The tariff is multiplied by this row's OWN measured input tokens (1,943,705 over 1,500 answered rows, mean 1,296) and divided by the 1,485 items of the common cost basis the scorer uses. The output term is zero twice over: the tariff bills no output, and the vendor's own usage meter reports 0 output tokens on all 1,500 rows. Gate chain: the pricing update made the cost axis conditional on a separate frozen-pricing change passing re-freeze, unchanged-row proof and independent review; the pricing review requested it; the scoped change adds this one key to pricing_v15.LAUNCH_LIST_PRICE_API and nothing else (r113/pricing/CANDIDATE.diff; unchanged-row proof r113/receipts/PRICING-EQUIV-R113.json and two rounds of independent review, r113/review-pricing/REVIEW.json (REVISE) and r113/review-pricing-r2/REVIEW.json (PASS); the findings about this row's text and the builder's gates are applied here, the two about the publication layer are on the build sheet). INSIDE THE OFFICIAL COST CAP on the common basis: USD 0.025716 per 1,000 decisions is 0.398x the official cap USD 0.064595. Over all 1,500 items the row cost USD 0.064790 (1.003x the cap) and was disclosed as ranked over the cap in v1.6.29; that disclosure is superseded. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.025716 per 1,000 decisions (0.398x the official cost cap USD 0.064595), against USD 0.064790 over all 1,500 items (v1.6.29)."},"latency":{"p50_s_raw":0.4660752070049057,"p50_s_adjusted":0.4660752070049057,"adjustment":"none (API)","p50_s_measured":0.4660752070049057,"latency_basis":"api-end-to-end","p95_s_raw":0.49626982034533285,"p95_s_adjusted":0.49626982034533285,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":74.38704101088614,"capability":81.47372650071756,"axes":{"calibration":90.69740052356492,"cost":57.693900899055016,"intelligence":72.25005247787018,"speed":86.40460211903752},"rank":8,"open_board_rank":null,"api_board_rank":7}},{"key":"surogate-rune-26b-a4b-v3-v16","name":"Surogate Rune 26B-A4B v3 (v1.6 pool)","board":"open","source_url":"https://huggingface.co/surogate/rune-26b-a4b-GGUF","axes":{"intelligence":56.055735926690254,"calibration":90.88042351612289,"speed":86.53673235812263,"cost":71.03441945513822},"capability":{"score":73.46807972140657,"rank":76,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":54,"api_board_rank":null},"composite_score":73.42639028852348,"rank":53,"open_board_rank":38,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009236753450580516,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 780,723 input + 1,477 output tokens over 1,477 decisions = 529.587000677 tokens/decision. GPU-time cost USD 0.00923675345058052/1,000; old list-price cost USD 0.0478728300609343/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00923675345058052/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11678496899548918,"p50_s_adjusted":0.3835699379909784,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11678496899548918,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.21436913508514407,"p95_s_adjusted":0.5787382701702881,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":65.0490738113794,"capability":73.46807972140657,"axes":{"calibration":90.88042351612289,"cost":49.597326937281316,"intelligence":56.055735926690254,"speed":86.95393114544444},"rank":48,"open_board_rank":27,"api_board_rank":null}},{"key":"swanone","name":"swanOne (blockbrain, Qwen3.8-Flash-Next NVFP4)","board":"open","source_url":"https://github.com/blockbrain-ai/swanone-recipe","axes":{"intelligence":56.15104294255194,"calibration":88.58151013990448,"speed":82.65355535838808,"cost":71.43288940092707},"capability":{"score":72.36627654122822,"rank":81,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":58,"api_board_rank":null},"composite_score":72.46902913443873,"rank":68,"open_board_rank":52,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008958535519807157,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 912,828 input + 1,624 output tokens over 1,624 decisions = 563.086206897 tokens/decision. GPU-time cost USD 0.00895853551980716/1,000; old list-price cost USD 0.0847829310344827/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00895853551980716/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.2801306745968759,"p50_s_adjusted":0.7102613491937518,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2801306745968759,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.30711693582125005,"p95_s_adjusted":0.7642338716425001,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":43.78640157551217,"capability":72.36627654122822,"axes":{"calibration":88.58151013990448,"cost":42.15074720483203,"intelligence":56.15104294255194,"speed":82.64878475169937},"rank":72,"open_board_rank":47,"api_board_rank":null}},{"key":"system-one-open","name":"system-one-open (Gemma 4 E2B LoRA on an L4)","board":"api","source_url":"https://github.com/mithalouni/system-one-open","axes":{"intelligence":27.94,"calibration":75.94,"speed":79.18,"cost":68.33},"capability":{"score":51.94,"rank":214,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":35},"composite_score":16.39,"rank":227,"open_board_rank":null,"api_board_rank":46,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011367487684729064,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: v1.5 measured-input proxy; no exact base-model floor applies"},"latency":{"p50_s_raw":0.514,"p50_s_adjusted":1.028,"adjustment":"x2 (demo)","p50_s_measured":0.514,"latency_basis":"demo-endpoint-x2","p95_s_raw":0.587,"p95_s_adjusted":1.174,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":16.39,"capability":51.94,"axes":{"calibration":75.94,"cost":68.33,"intelligence":27.94,"speed":79.32},"rank":152,"open_board_rank":null,"api_board_rank":38}},{"key":"system-one-sg","name":"system-one (Qwen3-8B, Sean Goedecke)","board":"open","source_url":"https://github.com/sgoedecke/system-one","axes":{"intelligence":29.271508779158296,"calibration":30.16473321941294,"speed":90.51961611379383,"cost":68.27616759897502},"capability":{"score":29.718120999285617,"rank":271,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":233,"api_board_rank":null},"composite_score":14.739752382089495,"rank":236,"open_board_rank":189,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011414610062569962,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 8B active parameters; throughput 30762.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 941,626 input + 1,624 output tokens over 1,624 decisions = 580.818965517 tokens/decision. GPU-time cost USD 0.01141461006257/1,000; old list-price cost USD 0.0682938189655172/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.01141461006257/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.053127793507883325,"p50_s_adjusted":0.2562555870157667,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.053127793507883325,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.09811513951281085,"p95_s_adjusted":0.3462302790256217,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":11.027723483913586,"capability":29.718120999285617,"axes":{"calibration":30.16473321941294,"cost":44.96855803039237,"intelligence":29.271508779158296,"speed":90.8867694864598},"rank":167,"open_board_rank":128,"api_board_rank":null}},{"key":"tacet-sonata","name":"Tacet Sonata (CodePawl, 144M packed encoder on mmBERT-small, option-marker softmax in one pass, in-process via the author's tacet package)","board":"open","source_url":"https://huggingface.co/codepawl/tacet-sonata","axes":{"intelligence":2.2096907014730043,"calibration":46.26085085939199,"speed":79.16975860375432,"cost":89.01467380421724},"capability":{"score":24.235270780432494,"rank":274,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":235,"api_board_rank":null},"composite_score":0.015686432078748844,"rank":305,"open_board_rank":252,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0023236865267433986,"basis":"ESTIMATE (size-class reference; self-hosted open weights, no public tariff). THE SAME CHOICE THE PUBLISHED quyet-1-0-tiny ROW CARRIES FOR THE IDENTICAL mmBERT-small BACKBONE: DeepInfra base-size encoders at USD 0.005 per 1M input, 0 output - deepinfra:thenlper/gte-base, deepinfra:BAAI/bge-base-en-v1.5, deepinfra:intfloat/e5-base-v2 and deepinfra:sentence-transformers/all-mpnet-base-v2 are all (0.005, 0.0) in the frozen 25 Sep snapshot - x this row's own measured usage.input_tokens (the package's packed length, capped at 4,096). Estimated, not charged. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, verified in the package source (tacet/decoding.py usage_of sets output_tokens 0; the readout is a softmax over option [MASK] markers from one encoder pass and nothing is generated), the published 02 Oct rule for classifier read-outs. PRICING REFERENCE NOTE: jhu-clsp/mmBERT-small still has no explicit market-listing decision (pricing_v15.reference() raises), so an UNLISTED_BASE_MODELS entry would require a published pricing update."},"latency":{"p50_s_raw":0.2647777907550335,"p50_s_adjusted":0.679555581510067,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.2647777907550335,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.8157773386687038,"p95_s_adjusted":1.7815546773374076,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":3,"n_over_limit_answered_ok":3,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.016405006835540154,"capability":24.310999978028352,"axes":{"calibration":46.37837260299999,"cost":89.01467380421724,"intelligence":2.2436273530567146,"speed":80.28211114890141},"rank":232,"open_board_rank":188,"api_board_rank":null}},{"key":"tde-general-large","name":"TDE general large v0.1 (tdelab, ModernBERT-large encoder)","board":"open","source_url":"https://huggingface.co/tdelab/tde-general-large-v0.1","axes":{"intelligence":5.640752949655241,"calibration":56.064346266900806,"speed":93.94749802780368,"cost":null},"capability":{"score":30.852549608278025,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.26","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-16","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"unpriced: answerdotai/ModernBERT-large is in the frozen module's UNLISTED_BASE_MODELS (no eligible base-model price); one listing decision would make the row rankable.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED. The frozen pricing_v15 module (sha256 3948e3f1..., and the later re-freezes, which only add or re-price hosted-API launch-list keys) has no market-listing decision for the declared base answerdotai/ModernBERT-large (the module lists it in UNLISTED_BASE_MODELS; the card declares no base_model); no price is invented here. Axes only: Intelligence, Calibration and Speed are computed; Cost, composite and rank are void. Tokens: the server reports no token field, so input is counted OFFLINE with the model's own tokenizer on the pod (the counter of the earlier measurement, 672,464 over the full draw, including the 1,024-token truncation) and output is 0 under the 9 Oct 2026 rule (encoder classifier, nothing generated)."},"latency":{"p50_s_raw":0.019978900207206607,"p50_s_adjusted":0.1899578004144132,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.019978900207206607,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.031062795454636215,"p95_s_adjusted":0.2121255909092724,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":30.852549608278025,"axes":{"calibration":56.064346266900806,"cost":null,"intelligence":5.640752949655241,"speed":93.9202043701742},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"tern-35b-a3b","name":"Tern-35B-A3B (matthewagi, Qwen3.6 35B-A3B MoE decision model, author server, self-hosted)","board":"open","source_url":"https://huggingface.co/matthewagi/Tern-35B-A3B","axes":{"intelligence":44.07459281825939,"calibration":88.71404111864382,"speed":90.72909335832355,"cost":71.33135335381719},"capability":{"score":66.3943169684516,"rank":125,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":98,"api_board_rank":null},"composite_score":52.678174120800755,"rank":139,"open_board_rank":109,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009028623746552559,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 842,725 input + 0 output tokens over 1,485 decisions = 567.491582492 tokens/decision. GPU-time cost USD 0.00902862374655256/1,000; old list-price cost USD 0.0851237373737374/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00902862374655256/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06874230350001653,"p50_s_adjusted":0.2874846070000331,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06874230350001653,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07204358164991617,"p95_s_adjusted":0.2940871632998323,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"torchcast-decision-12b","name":"torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout)","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-12b","axes":{"intelligence":60.54320091058725,"calibration":82.93469131589123,"speed":91.34133469499314,"cost":56.36034169226667},"capability":{"score":71.73894611323924,"rank":85,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":62,"api_board_rank":null},"composite_score":69.84980331350847,"rank":100,"open_board_rank":78,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02848756157635468,"basis":"v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (v1.5-M2 base-model reference; no public tariff, customer asked for the base-model reference): OpenRouter google/gemma-3-12b-it list price $0.05/M input, $0.15/M output (gemma-4-12b-it is not listed on OpenRouter, read 2026-10-04; same reference as the Cygnet and Jev-Omni Gemma-4-12B rows) x the shim's own measured usage.prompt_tokens/completion_tokens; self-hosted on 1x H100 80GB (Lium), vLLM 0.30.0 BF16; estimated, not charged"},"latency":{"p50_s_raw":0.04482428729534149,"p50_s_adjusted":0.23964857459068298,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04482428729534149,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07820102982223033,"p95_s_adjusted":0.30640205964446066,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":69.9087143535159,"capability":71.73894611323924,"axes":{"calibration":82.93469131589123,"cost":56.36034169226667,"intelligence":60.54320091058725,"speed":91.74573681567192},"rank":27,"open_board_rank":11,"api_board_rank":null}},{"key":"torchcast-decision-27b","name":"torchcast-decision-27b (Torchcast AI, Qwen3.8-27B fine-tune)","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-27b","axes":{"intelligence":70.39456337588172,"calibration":90.04368712409786,"speed":87.62649746267333,"cost":57.71080894844077},"capability":{"score":80.21912524998979,"rank":35,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":26,"api_board_rank":null},"composite_score":74.00389040716802,"rank":43,"open_board_rank":30,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.025682642218820436,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 738,068 input + 0 output tokens over 1,477 decisions = 499.707515234 tokens/decision. GPU-time cost USD 0.0256826422188204/1,000; old list-price cost USD 0.209877156398104/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0256826422188204/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.09896963900246192,"p50_s_adjusted":0.34793927800492386,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09896963900246192,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.17320858579478227,"p95_s_adjusted":0.49641717158956455,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":23,"n_over_limit_answered_ok":23,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.273291266043532,"capability":81.56565836114741,"axes":{"calibration":91.379440941337,"cost":30.341044859222123,"intelligence":71.75187578095783,"speed":87.72436167205926},"rank":123,"open_board_rank":89,"api_board_rank":null}},{"key":"torchcast-decision-31b-v1.0","name":"Torchcast Decision 31B v1.0","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-31b","axes":{"intelligence":85.47425118723262,"calibration":93.42911792902787,"speed":87.26135463003386,"cost":59.41752684953559},"capability":{"score":89.45168455813024,"rank":1,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":1,"api_board_rank":null},"composite_score":78.90696704238127,"rank":3,"open_board_rank":2,"api_board_rank":null,"measured_in":"v1.7.14","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.14","measured_on":"2026-10-27","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02252937298362935,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 560,794 input + 0 output tokens over 1,489 decisions = 376.624580255 tokens/decision. GPU-time cost USD 0.0225293729836294/1,000; old list-price cost USD 0.0338962122229684/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0225293729836294/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.09475059950000286,"p50_s_adjusted":0.33950119900000575,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09475059950000286,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.20168985064995013,"p95_s_adjusted":0.5533797012999002,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.14","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"torchcast-decision-4b-v1.1","name":"Torchcast Decision 4B (v1.1)","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-4b","axes":{"intelligence":54.20841827353793,"calibration":83.11679203980711,"speed":93.5282853940162,"cost":79.35029829231833},"capability":{"score":68.66260515667253,"rank":115,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":88,"api_board_rank":null},"composite_score":74.38693697091249,"rank":38,"open_board_rank":26,"api_board_rank":null,"measured_in":"v1.6.12","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004878916553854475,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 737,326 input + 0 output tokens over 1,485 decisions = 496.515824916 tokens/decision. GPU-time cost USD 0.00487891655385447/1,000; old list-price cost USD 0.0148954747474747/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00487891655385447/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02236862655263394,"p50_s_adjusted":0.19473725310526788,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02236862655263394,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.03894425290636718,"p95_s_adjusted":0.22788850581273434,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.68474930403303,"capability":68.66260515667253,"axes":{"calibration":83.11679203980711,"cost":64.8083695127977,"intelligence":54.20841827353793,"speed":93.62765623388151},"rank":23,"open_board_rank":9,"api_board_rank":null}},{"key":"torchcast-decision-4b-v1.2","name":"Torchcast Decision 4B (v1.2)","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-4b","axes":{"intelligence":60.262524104890716,"calibration":84.9508191193537,"speed":90.89431110061412,"cost":79.35029829231833},"capability":{"score":72.6066716121222,"rank":80,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":57,"api_board_rank":null},"composite_score":76.9679054341264,"rank":15,"open_board_rank":10,"api_board_rank":null,"measured_in":"v1.6.29","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-18","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004878916553854475,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 737,326 input + 0 output tokens over 1,485 decisions = 496.515824916 tokens/decision. GPU-time cost USD 0.00487891655385447/1,000; old list-price cost USD 0.0148954747474747/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00487891655385447/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06279662949995668,"p50_s_adjusted":0.2755932589999134,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06279662949995668,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.0726625294999678,"p95_s_adjusted":0.2953250589999356,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":73.03338420137015,"capability":72.6066716121222,"axes":{"calibration":84.9508191193537,"cost":64.8083695127977,"intelligence":60.262524104890716,"speed":91.12622136196717},"rank":14,"open_board_rank":4,"api_board_rank":null}},{"key":"torchcast-decision-4b-v1.3","name":"Torchcast Decision 4B (v1.3)","board":"open","source_url":"https://huggingface.co/torchcast-ai/torchcast-decision-4b","axes":{"intelligence":67.14748494108278,"calibration":87.84716845590987,"speed":92.44891117449383,"cost":79.35029829231833},"capability":{"score":77.49732669849632,"rank":50,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":37,"api_board_rank":null},"composite_score":80.49078635844445,"rank":1,"open_board_rank":1,"api_board_rank":null,"measured_in":"v1.6.42","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004878916553854475,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 737,326 input + 0 output tokens over 1,485 decisions = 496.515824916 tokens/decision. GPU-time cost USD 0.00487891655385447/1,000; old list-price cost USD 0.0148954747474747/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00487891655385447/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.033565528499991615,"p50_s_adjusted":0.21713105699998322,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.033565528499991615,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05602583630003721,"p95_s_adjusted":0.26205167260007445,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":76.18936170949911,"capability":77.49732669849632,"axes":{"calibration":87.84716845590987,"cost":64.8083695127977,"intelligence":67.14748494108278,"speed":92.63835078333227},"rank":3,"open_board_rank":1,"api_board_rank":null}},{"key":"trio-spark-v1.3","name":"Trio-Spark v1.3 (MachineFi)","board":"api","source_url":"https://github.com/machinefi/trio-spark","axes":{"intelligence":28.409112519894506,"calibration":85.64268754215436,"speed":76.18058448046628,"cost":null},"capability":{"score":57.025900031024435,"rank":null,"eligible":false,"within_caps":false,"reasons":["no cost reported","latency 2.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":null,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.22","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-15","status":"anchor-runs-pending"},"ranked":false,"listing":"listed","not_ranked_because":"Measured, but the frozen pricing path gives it no eligible price: MachineFi publishes a usage tariff for Trio-Spark (USD 0.042 per 1M billed input tokens, output not billed), yet `trio-spark-v1.3` is in neither frozen tariff set (`AGE_QUALIFIED_API`, `LAUNCH_LIST_PRICE_API`) and the system declares no base model and releases no weights, so there is no base-model market reference either — the frozen module (unchanged on this point since the 32117d75... module it was first checked against) returns \"no eligible 30-day list price or base-model estimate\" (verified by executing it; measurement receipts on file). Without a cost axis there is no composite, so the row shows Intelligence, Calibration and Capability and is not ranked, the same outcome as `sage-1.4.0` and as `subq-snap-preview` before the pricing update. The pricing question is open and can be decided in a later release without re-measuring anything.","price":{"kind":"unpriced","usd_per_1000_decisions":null,"basis":"UNPRICED under the frozen pricing_v15 path (sha256 dcfcecb06525ce4a5ece37400ea261b97d2fe5d41155f3cd02b03cac86572835, the module in force since v1.6.21; also under its predecessor 32117d754104b3ead8bebdc9da5f8e0116a7550f4316cfc6874fbe6234c2d72b, against which it was first checked): MachineFi publishes a usage tariff for Trio-Spark of USD 0.042 per 1M billed input tokens with no output charge (docs/visual.md and the README of machinefi/trio-spark), but `trio-spark-v1.3` is in neither AGE_QUALIFIED_API nor LAUNCH_LIST_PRICE_API, and the system declares no base model and releases no weights, so there is no base-model market reference to fall back on. The frozen path therefore returns no eligible price (the same outcome as sage-1.4.0, and as subq-snap-preview before the pricing update)."},"latency":{"p50_s_raw":1.3097197005117778,"p50_s_adjusted":1.3097197005117778,"adjustment":"none (API)","p50_s_measured":1.3097197005117778,"latency_basis":"api-end-to-end","p95_s_raw":1.8397685555450152,"p95_s_adjusted":1.8397685555450152,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":null,"capability":57.025900031024435,"axes":{"calibration":85.64268754215436,"cost":null,"intelligence":28.409112519894506,"speed":76.46737050345754},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"turion-jev-31b","name":"Turion Jev 31B","board":"open","source_url":"https://huggingface.co/turion-ai/turion-jev-31b","axes":{"intelligence":70.85736825150448,"calibration":90.90010349317609,"speed":85.85779594425273,"cost":46.586791013615496},"capability":{"score":80.87873587234029,"rank":31,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":23,"api_board_rank":null},"composite_score":59.63832789350321,"rank":131,"open_board_rank":101,"api_board_rank":null,"measured_in":"v1.7.1","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.1","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.06031707871288843,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10664.9184474 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=native). The row's own scored cost basis: 1,580,092 input + 0 output tokens over 1,485 decisions = 1064.03501684 tokens/decision. GPU-time cost USD 0.0603170787128884/1,000; old list-price cost USD 0.0957631515151515/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0603170787128884/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.12411595150007315,"p50_s_adjusted":0.3982319030001463,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.12411595150007315,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.25087748864994525,"p95_s_adjusted":0.6517549772998905,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"typecastlm","name":"typecastlm (Mikhail Gribov, Qwen3.5-4B computed head)","board":"open","source_url":"https://github.com/mihail-gribov/typecastlm","axes":{"intelligence":29.78900656204624,"calibration":83.29951051808085,"speed":92.66059621716312,"cost":78.50093844783655},"capability":{"score":56.544258540063545,"rank":190,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":156,"api_board_rank":null},"composite_score":20.54575277268221,"rank":213,"open_board_rank":171,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.005207573221225232,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 860,659 input + 0 output tokens over 1,624 decisions = 529.962438424 tokens/decision. GPU-time cost USD 0.00520757322122523/1,000; old list-price cost USD 0.0158988731527094/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00520757322122523/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03045072252280079,"p50_s_adjusted":0.21090144504560157,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03045072252280079,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05347860935726202,"p95_s_adjusted":0.25695721871452404,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":19.72256430004839,"capability":56.544258540063545,"axes":{"calibration":83.29951051808085,"cost":63.95900966831593,"intelligence":29.78900656204624,"speed":92.76354061232337},"rank":133,"open_board_rank":98,"api_board_rank":null}},{"key":"typecastlm-1.4.0","name":"TypeCastLM 1.4.0 (Mikhail Gribov, Qwen3.5-4B computed head)","board":"open","source_url":"https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b","axes":{"intelligence":35.43137209354293,"calibration":77.35692434429644,"speed":92.811325470547,"cost":79.34178886317649},"capability":{"score":56.39414821891969,"rank":193,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":159,"api_board_rank":null},"composite_score":31.127309270665172,"rank":188,"open_board_rank":149,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004882104126432508,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 733,833 input + 0 output tokens over 1,477 decisions = 496.840216655 tokens/decision. GPU-time cost USD 0.00488210412643251/1,000; old list-price cost USD 0.0149052064996615/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00488210412643251/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.03054074199462775,"p50_s_adjusted":0.2110814839892555,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.03054074199462775,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.048990187999152106,"p95_s_adjusted":0.2479803759983042,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":29.827386072784904,"capability":56.39414821891969,"axes":{"calibration":77.35692434429644,"cost":64.79986008365586,"intelligence":35.43137209354293,"speed":92.95077486551261},"rank":102,"open_board_rank":70,"api_board_rank":null}},{"key":"umpire-35b-a3b","name":"Umpire 35B-A3B","board":"open","source_url":"https://huggingface.co/ezoushen/umpire-35b-a3b-bf16","axes":{"intelligence":55.30307844451234,"calibration":88.87842513440546,"speed":79.61785431820445,"cost":75.39758639297672},"capability":{"score":72.09075178945889,"rank":83,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":60,"api_board_rank":null},"composite_score":72.520889614644,"rank":67,"open_board_rank":51,"api_board_rank":null,"measured_in":"v1.7.22","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.22","measured_on":"2026-10-29","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.006608158535893478,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe qwen, 3B active parameters; throughput 37999.6666667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 618,462 input + 0 output tokens over 1,489 decisions = 415.353928811 tokens/decision. GPU-time cost USD 0.00660815853589348/1,000; old list-price cost USD 0.129798102753526/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00660815853589348/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.40499839199992493,"p50_s_adjusted":0.9599967839998499,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.40499839199992493,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.49374133670019077,"p95_s_adjusted":1.1374826734003816,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.22","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"uniphore-4b-decider","name":"Uniphore-4B-Decider","board":"open","source_url":"https://huggingface.co/Uniphore/uniphore-4b-decider","axes":{"intelligence":34.15161183674326,"calibration":79.97905975708014,"speed":87.30790070278863,"cost":79.08243125416328},"capability":{"score":57.0653357969117,"rank":185,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":151,"api_board_rank":null},"composite_score":28.32484800212957,"rank":194,"open_board_rank":154,"api_board_rank":null,"measured_in":"v1.6.11","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004980263157580418,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 752,642 input + 0 output tokens over 1,485 decisions = 506.82962963 tokens/decision. GPU-time cost USD 0.00498026315758042/1,000; old list-price cost USD 0.0152048888888889/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00498026315758042/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.10890222905436531,"p50_s_adjusted":0.36780445810873064,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.10890222905436531,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.17767539023188872,"p95_s_adjusted":0.5053507804637775,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":27.171470959871687,"capability":57.0653357969117,"axes":{"calibration":79.97905975708014,"cost":64.54050247464266,"intelligence":34.15161183674326,"speed":87.70990379280212},"rank":107,"open_board_rank":74,"api_board_rank":null}},{"key":"vansa-3.4","name":"Vansa-3.4 (Vansa, hosted System One API)","board":"api","source_url":"https://vansa.org","axes":{"intelligence":40.85,"calibration":86.52,"speed":95.05,"cost":61.92},"capability":{"score":63.685,"rank":142,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":30},"composite_score":42.57,"rank":166,"open_board_rank":null,"api_board_rank":35,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.018590751269035533,"basis":"measured: token usage of the 6 Oct 2026 A4 u P run x tariff USD 0.034/M input, 0/M output (spec vansa-3.4.json), per 1,000 decisions on the v1.6.1 common cost basis (591 of 600 items; the 9 A4 u P items outside the Jev reference's accepted input range are excluded, cost-basis-a4.json)"},"latency":{"p50_s_raw":0.158,"p50_s_adjusted":0.158,"adjustment":"none (API)","p50_s_measured":0.158,"latency_basis":"api-end-to-end","p95_s_raw":0.198,"p95_s_adjusted":0.198,"n":600},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":600,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":42.6,"capability":63.68,"axes":{"calibration":86.52,"cost":61.92,"intelligence":40.85,"speed":95.49},"rank":75,"open_board_rank":null,"api_board_rank":27}},{"key":"vega-u-qwen3.5-0.8b","name":"Vega-U (Qwen3.5-0.8B reader)","board":"open","source_url":"https://huggingface.co/nandakishorm/vega-U","axes":{"intelligence":5.342898619124496,"calibration":66.9335611044975,"speed":73.05648557202497,"cost":85.77193998006749},"capability":{"score":36.138229861811,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 2.6× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":0.2008099765046913,"rank":287,"open_board_rank":234,"api_board_rank":null,"measured_in":"v1.6.29","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-18","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0029803595972957087,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 732,071 input + 0 output tokens over 1,485 decisions = 492.977104377 tokens/decision. GPU-time cost USD 0.00298035959729571/1,000; old list-price cost USD 0.00492977104377104/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00298035959729571/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.7303033574862638,"p50_s_adjusted":1.6106067149725276,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.7303033574862638,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.460790375790384,"p95_s_adjusted":3.071580751580768,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.21762292373052544,"capability":36.52005519443276,"axes":{"calibration":67.53669192481355,"cost":79.21519751338194,"intelligence":5.503418464051958,"speed":73.8348503635803},"rank":213,"open_board_rank":169,"api_board_rank":null}},{"key":"vega-u-qwen3.5-4b","name":"Vega-U (Qwen3.5-4B reader)","board":"open","source_url":"https://huggingface.co/nandakishorm/vega-U","axes":{"intelligence":8.45673307172142,"calibration":76.81388603317517,"speed":88.8199586422866,"cost":79.44348865131269},"capability":{"score":42.63530955244829,"rank":244,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":207,"api_board_rank":null},"composite_score":0.7376917704778548,"rank":275,"open_board_rank":223,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004844144002105987,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 732,071 input + 0 output tokens over 1,485 decisions = 492.977104377 tokens/decision. GPU-time cost USD 0.00484414400210599/1,000; old list-price cost USD 0.0147893131313131/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00484414400210599/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.09011760349994802,"p50_s_adjusted":0.33023520699989606,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.09011760349994802,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.12367845209996063,"p95_s_adjusted":0.39735690419992126,"n":1500},"method":"v1.7","rescored_in":null,"remeasured_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.7843042690267673,"capability":44.31251291226081,"axes":{"calibration":79.8233873834119,"cost":64.90155987179207,"intelligence":8.801638441109716,"speed":60.651933418279626},"rank":199,"open_board_rank":156,"api_board_rank":null}},{"key":"vegaml-0.8b","name":"Vega 0.8B","board":"open","source_url":"https://huggingface.co/nandakishorm/vega-08b-public-intents","axes":{"intelligence":3.0532964296406795,"calibration":62.93608607777928,"speed":89.16338093060068,"cost":68.03119990679988},"capability":{"score":32.99469125370998,"rank":263,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":225,"api_board_rank":null},"composite_score":0.040388417880414994,"rank":301,"open_board_rank":248,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011631257480158733,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,885,865 input + 0 output tokens over 1,500 decisions = 1923.91 tokens/decision. GPU-time cost USD 0.0116312574801587/1,000; old list-price cost USD 0.0192391/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0116312574801587/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.08674963191151619,"p50_s_adjusted":0.3234992638230324,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.08674963191151619,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.11239526560530065,"p95_s_adjusted":0.3747905312106013,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.04022166962784317,"capability":32.99469125370998,"axes":{"calibration":62.93608607777928,"cost":61.47445744011434,"intelligence":3.0532964296406795,"speed":89.45598874418484},"rank":227,"open_board_rank":183,"api_board_rank":null}},{"key":"vegaml-4b","name":"Vega 4B","board":"open","source_url":"https://huggingface.co/nandakishorm/vega-08b-public-intents","axes":{"intelligence":15.69626515886651,"calibration":57.46108428162757,"speed":83.93992701495202,"cost":61.702748578045096},"capability":{"score":36.57867472024704,"rank":258,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":221,"api_board_rank":null},"composite_score":3.608782255985968,"rank":264,"open_board_rank":215,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.018904928858445722,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 2,885,865 input + 0 output tokens over 1,500 decisions = 1923.91 tokens/decision. GPU-time cost USD 0.0189049288584457/1,000; old list-price cost USD 0.0577173/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0189049288584457/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1971108582802117,"p50_s_adjusted":0.5442217165604234,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1971108582802117,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.29585269126109776,"p95_s_adjusted":0.7417053825221955,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":29,"n_over_limit_answered_ok":29,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.603902446043355,"capability":37.687052549362136,"axes":{"calibration":58.68489087403218,"cost":47.16081979852447,"intelligence":16.689214224692094,"speed":84.39386641436894},"rank":188,"open_board_rank":148,"api_board_rank":null}},{"key":"vera-large","name":"Vera Large","board":"open","source_url":"https://huggingface.co/scar-ai/vera-large","axes":{"intelligence":67.78675651753129,"calibration":91.31679956749362,"speed":88.01635603332511,"cost":55.930081512098035},"capability":{"score":79.55177804251245,"rank":41,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":31,"api_board_rank":null},"composite_score":72.80136805547387,"rank":65,"open_board_rank":49,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.029444032127162855,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 766,685 input + 84,061 output tokens over 1,485 decisions = 572.892929293 tokens/decision. GPU-time cost USD 0.0294440321271629/1,000; old list-price cost USD 0.216840202020202/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0294440321271629/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0924675480000019,"p50_s_adjusted":0.3349350960000038,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0924675480000019,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.16070768570001512,"p95_s_adjusted":0.4714153714000302,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"verdict-small","name":"verdict-small (Manavarya09, multilingual-e5-small 118M)","board":"open","source_url":"https://github.com/Manavarya09/verdict","axes":{"intelligence":2.3417095945172277,"calibration":72.92000413192591,"speed":88.74744771425449,"cost":100},"capability":{"score":37.63085686322157,"rank":256,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":219,"api_board_rank":null},"composite_score":0.018990012754194416,"rank":304,"open_board_rank":251,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.000866152709359606,"basis":"v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; no exact base-model floor applies"},"latency":{"p50_s_raw":0.0651877773925662,"p50_s_adjusted":0.28037555478513243,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0651877773925662,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.1629496591165662,"p95_s_adjusted":0.47589931823313236,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.01899384266854915,"capability":37.63085686322157,"axes":{"calibration":72.92000413192591,"cost":100,"intelligence":2.3417095945172277,"speed":89.48731594094681},"rank":231,"open_board_rank":187,"api_board_rank":null}},{"key":"verdin-4b","name":"Verdin-4B (Victor Bennett, Qwen3.5-4B + merged LoRA r16 + pointer head)","board":"open","source_url":"https://huggingface.co/victorbennett/Verdin-4B","axes":{"intelligence":35.44491489257096,"calibration":83.75263461013449,"speed":90.21057479487555,"cost":76.72117336001418},"capability":{"score":59.59877475135272,"rank":170,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":138,"api_board_rank":null},"composite_score":31.275579231606862,"rank":187,"open_board_rank":148,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-22","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.00596981520777629,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 902,188 input + 0 output tokens over 1,485 decisions = 607.534006734 tokens/decision. GPU-time cost USD 0.00596981520777629/1,000; old list-price cost USD 0.0182260202020202/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00596981520777629/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.0580428564999238,"p50_s_adjusted":0.2660857129998476,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.0580428564999238,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.10401563704999713,"p95_s_adjusted":0.35803127409999425,"n":1500},"method":"v1.7","rescored_in":null,"remeasured_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":21.77748190312391,"capability":59.540149828182294,"axes":{"calibration":83.57547250683973,"cost":62.17924458049355,"intelligence":35.50482714952485,"speed":45.8869739530126},"rank":117,"open_board_rank":84,"api_board_rank":null}},{"key":"visionjev-1.0.0-27b","name":"VisionJEV 1.0.0","board":"open","source_url":"https://huggingface.co/taoszhang/VisionJEV","axes":{"intelligence":75.89673617998416,"calibration":92.46253951833408,"speed":84.80563168345245,"cost":61.42265544156546},"capability":{"score":84.17963784915912,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":76.82955864559965,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.7.21","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.7.21","measured_on":"2026-10-29","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Measured on the 27 Oct 2026 draw and shown on the reference scale by anchor equating with four strong anchor systems (Intelligence -2.9, Calibration -1.9 points). Not ranked yet: its top-five places lie within the uncertainty of that shift (95% interval -5.9 to +1.1 Intelligence points). Ranking follows when more anchor runs narrow the interval.","price":{"kind":"estimate","usd_per_1000_decisions":0.019315745975156336,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 559,606 input + 0 output tokens over 1,489 decisions = 375.826729349 tokens/decision. GPU-time cost USD 0.0193157459751563/1,000; old list-price cost USD 0.157847226326394/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0193157459751563/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15154025199990429,"p50_s_adjusted":0.4530805039998086,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15154025199990429,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2899483975500857,"p95_s_adjusted":0.7298967951001715,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.21","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"vllm-sr-d3","name":"vLLM-SR d3","board":"open","source_url":"https://huggingface.co/vllm-sr/d3","axes":{"intelligence":72.92342726693953,"calibration":91.37586990638505,"speed":87.00758628687251,"cost":57.27607555734765},"capability":{"score":82.1496485866623,"rank":17,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":12,"api_board_rank":null},"composite_score":74.61382220232323,"rank":34,"open_board_rank":23,"api_board_rank":null,"measured_in":"v1.6.16","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-12","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.02655405281741362,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11823.4995401 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=trunk-equivalent). The row's own scored cost basis: 771,190 input + 0 output tokens over 1,485 decisions = 519.31986532 tokens/decision. GPU-time cost USD 0.0265540528174136/1,000; old list-price cost USD 0.218114343434343/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0265540528174136/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.11013474199990014,"p50_s_adjusted":0.3702694839998003,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.11013474199990014,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.19396356639996384,"p95_s_adjusted":0.5379271327999277,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.53757421097071,"capability":82.96592796157353,"axes":{"calibration":92.20028306358832,"cost":29.839473214267542,"intelligence":73.73157285955872,"speed":87.36893361347018},"rank":128,"open_board_rank":94,"api_board_rank":null}},{"key":"vllm-sr-d3-mini","name":"vLLM-SR d3-mini","board":"open","source_url":"https://huggingface.co/vllm-sr/d3-mini","axes":{"intelligence":50.78521313506569,"calibration":86.75010672241689,"speed":93.06198600681341,"cost":78.76524306413529},"capability":{"score":68.7676599287413,"rank":114,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":87,"api_board_rank":null},"composite_score":73.18017064643449,"rank":59,"open_board_rank":44,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.0051029960386139005,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 771,190 input + 0 output tokens over 1,485 decisions = 519.31986532 tokens/decision. GPU-time cost USD 0.0051029960386139/1,000; old list-price cost USD 0.015579595959596/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0051029960386139/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02826550850002718,"p50_s_adjusted":0.20653101700005436,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02826550850002718,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.044615130949964055,"p95_s_adjusted":0.2392302618999281,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":15,"n_over_limit_answered_ok":15,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"vllm-sr-d3-turbo","name":"vLLM-SR d3-turbo","board":"open","source_url":"https://huggingface.co/vllm-sr/d3-turbo","axes":{"intelligence":74.82897422466625,"calibration":90.35330442455489,"speed":85.64377521409548,"cost":57.25589630946454},"capability":{"score":82.59113932461057,"rank":15,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":10,"api_board_rank":null},"composite_score":74.66420728011205,"rank":33,"open_board_rank":22,"api_board_rank":null,"measured_in":"v1.6.31","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.6.43","measured_on":"2026-10-19","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.026595211985503357,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11805.2013064 tokens/s (direct measurement, 29 Oct 2026, stock vLLM 0.32.0 EFF, fresh 200 public decisions; load_mode=trunk-equivalent). The row's own scored cost basis: 771,190 input + 0 output tokens over 1,485 decisions = 519.31986532 tokens/decision. GPU-time cost USD 0.0265952119855034/1,000; old list-price cost USD 0.218114343434343/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0265952119855034/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.15096561499996142,"p50_s_adjusted":0.45193122999992286,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.15096561499996142,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.22666166885008465,"p95_s_adjusted":0.6033233377001693,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":20.47854238643572,"capability":82.59113932461057,"axes":{"calibration":90.35330442455489,"cost":29.839473214267542,"intelligence":74.82897422466625,"speed":86.03728492328564},"rank":129,"open_board_rank":95,"api_board_rank":null}},{"key":"wald-26b-a4b","name":"Wald 26B-A4B (Gemma 4 26B-A4B)","board":"open","source_url":"https://huggingface.co/org2ai/Wald-26B-A4B","axes":{"intelligence":65.50854607947286,"calibration":88.05577650422971,"speed":92.01872463451872,"cost":70.95910393447916},"capability":{"score":76.78216129185128,"rank":58,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":43,"api_board_rank":null},"composite_score":77.54703220970119,"rank":10,"open_board_rank":5,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-23","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009290302791701198,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 790,996 input + 0 output tokens over 1,485 decisions = 532.657239057 tokens/decision. GPU-time cost USD 0.0092903027917012/1,000; old list-price cost USD 0.0479391515151515/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0092903027917012/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.043133240500026204,"p50_s_adjusted":0.2362664810000524,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.043133240500026204,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.057952169350022585,"p95_s_adjusted":0.26590433870004515,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"wald-4b-v2","name":"Wald 4B","board":"open","source_url":"https://huggingface.co/org2ai/Wald-4B","axes":{"intelligence":41.00304543028662,"calibration":78.10292780844634,"speed":82.31428682922845,"cost":70.53846200925558},"capability":{"score":59.55298661936648,"rank":171,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":139,"api_board_rank":null},"composite_score":42.35067865852022,"rank":167,"open_board_rank":132,"api_board_rank":null,"measured_in":"v1.6.2","draw":{"id":"v1.6-fastlane-20261009","label":"Fast-lane draw","release":"v1.6.4","measured_on":"2026-10-09","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.009595138907940727,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 1,394,108 input + 50,098 output tokens over 1,479 decisions = 976.47464503 tokens/decision. GPU-time cost USD 0.00959513890794073/1,000; old list-price cost USD 0.033358985801217/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00959513890794073/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04425571905449033,"p50_s_adjusted":0.23851143810898065,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04425571905449033,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.1553597365040331,"p95_s_adjusted":2.460719473008066,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":8,"n_over_limit_answered_ok":8,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":39.9690949735785,"capability":59.75558427184466,"axes":{"calibration":78.39957889891498,"cost":54.30361484501323,"intelligence":41.11158964477434,"speed":82.20577865115169},"rank":83,"open_board_rank":54,"api_board_rank":null}},{"key":"wald-4b-v2-1","name":"Wald 4B v2.1","board":"open","source_url":"https://huggingface.co/org2ai/Wald-4B","axes":{"intelligence":51.951877017848766,"calibration":86.58998638479137,"speed":92.782872129877,"cost":60.82893594182271},"capability":{"score":69.27093170132007,"rank":112,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":85,"api_board_rank":null},"composite_score":68.94809240435124,"rank":106,"open_board_rank":82,"api_board_rank":null,"measured_in":"v1.6.5","draw":{"id":"v1.6-regular-20261010-a2","label":"Regular draw","release":"v1.6.9","measured_on":"2026-10-10","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020216321487157324,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 3,086,051 input + 0 output tokens over 1,500 decisions = 2057.36733333 tokens/decision. GPU-time cost USD 0.0202163214871573/1,000; old list-price cost USD 0.06172102/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0202163214871573/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.02897872356697917,"p50_s_adjusted":0.20795744713395833,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.02897872356697917,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.05168007561005651,"p95_s_adjusted":0.253360151220113,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":54.28713266716134,"capability":69.27093170132007,"axes":{"calibration":86.58998638479137,"cost":46.287007162302075,"intelligence":51.951877017848766,"speed":93.07746270925374},"rank":62,"open_board_rank":38,"api_board_rank":null}},{"key":"watersheep","name":"WaterSheep (Samrat Dutta, ModernBERT-base encoder with calibrated heads, served by the author's own `watersheep --serve`)","board":"open","source_url":"https://huggingface.co/samratduttaofficial/WaterSheep","axes":{"intelligence":5.305239654184986,"calibration":65.7809648808148,"speed":77.39270161387154,"cost":68.67136467637603},"capability":{"score":35.5431022674999,"rank":259,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":222,"api_board_rank":null},"composite_score":0.19479747591837096,"rank":288,"open_board_rank":235,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.011073574813811781,"basis":"ESTIMATE, labelled, and RUN 48'S RATE FOR THIS EXACT ROW AND BASE, CARRIED UNCHANGED: answerdotai/ModernBERT-base IS in the frozen pricing_v15.UNLISTED_BASE_MODELS, so no exact base-model floor applies and the module takes the estimate rate from the row's own registry entry; USD 0.03/M input, 0 output is the rate run 47 (quyet-10-small-en) and run 48 (watersheep on v1.5) already assigned to this base. Output 0 is literal: watersheep/infer.py returns output_tokens 0 and the serving path has no generation step (published 02 Oct rule). x this row's own measured input tokens. Estimated, not charged. No new decision."},"latency":{"p50_s_raw":0.462554931640625,"p50_s_adjusted":1.07510986328125,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.462554931640625,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.7727093957364559,"p95_s_adjusted":1.6954187914729117,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":1,"n_over_limit_answered_ok":1,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.19483758736149703,"capability":35.57380209564805,"axes":{"calibration":65.84236453711111,"cost":68.67136467637603,"intelligence":5.305239654184986,"speed":77.5933742155104},"rank":215,"open_board_rank":171,"api_board_rank":null}},{"key":"watt-flash-0.1","name":"watt-flash-0.1 (Zaitgeist Labs, 140M mmBERT-small encoder with an option-marker scorer, one forward pass per request, self-hosted)","board":"open","source_url":"https://huggingface.co/zaitlabs/watt-flash-0.1","axes":{"intelligence":3.759262959387639,"calibration":79.37888213235459,"speed":73.82860858123392,"cost":89.00253166733508},"capability":{"score":41.569072545871116,"rank":249,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":212,"api_board_rank":null},"composite_score":0.07452931157363736,"rank":297,"open_board_rank":244,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"mixed","usd_per_1000_decisions":0.0023258530805687202,"basis":"ESTIMATE (size-class reference; self-hosted open weights, no public tariff). THE SAME CHOICE THE PUBLISHED quyet-1-0-tiny ROW AND RUN 59'S tacet-sonata ROW CARRY FOR THE IDENTICAL mmBERT-small BACKBONE, carried unchanged and not a new judgement: DeepInfra base-size encoders at USD 0.005 per 1M input, 0 output - deepinfra:thenlper/gte-base, deepinfra:BAAI/bge-base-en-v1.5, deepinfra:intfloat/e5-base-v2 and deepinfra:sentence-transformers/all-mpnet-base-v2 are all (0.005, 0.0) in the frozen 25 Sep snapshot - x this row's own measured input tokens. Estimated, not charged. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, verified in the author's source: watt_flash/serve.py:answer() formats a softmax over the option markers of one encoder pass and the package contains no generation path at all (the published 02 Oct rule for classifier read-outs). THE SERVER REPORTS NO usage BLOCK OF ANY KIND, so the input counts come EXACTLY from the author's own unmodified watt_flash.layout.Layout.pack_request - the very number their own WattFlash.decide_batch returns as len(p) and only serve.py drops - under PREREG-R61-AMENDMENT-1, cross-checked three ways (independent reassembly of their documented layout; identity of their packer's Overflow set with the server's own 422 set), with no cost number at all on any disagreement; the counts sit in cost_estimate and never in usage. PRICING REFERENCE NOTE: jhu-clsp/mmBERT-small still has no explicit market-listing decision (pricing_v15.reference() raises), so an UNLISTED_BASE_MODELS entry would require a published pricing update."},"latency":{"p50_s_raw":0.49060964956879616,"p50_s_adjusted":1.1312192991375922,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.49060964956879616,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-CPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":1.755468840524554,"p95_s_adjusted":3.660937681049108,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":0.07461802002995101,"capability":41.569072545871116,"axes":{"calibration":79.37888213235459,"cost":89.00253166733508,"intelligence":3.759262959387639,"speed":75.84832436427908},"rank":223,"open_board_rank":179,"api_board_rank":null}},{"key":"wemm-jev","name":"WeMM-JEV (Rover00)","board":"open","source_url":"https://huggingface.co/Rover00/WeMM-JEV","axes":{"intelligence":78.32318029057845,"calibration":94.25822063430104,"speed":88.0997638122694,"cost":61.25046269001649},"capability":{"score":86.29070046243974,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":78.34850435521572,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.8.1","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.1","measured_on":"2026-10-29","status":"equated"},"ranked":false,"listing":"unranked","not_ranked_because":"Measured on the 27 Oct 2026 draw and shown on the reference scale by anchor equating with four strong anchor systems (Intelligence -2.9, Calibration -1.9 points). Not ranked yet: its top-five places lie within the uncertainty of that shift (95% interval -5.9 to +1.1 Intelligence points). Ranking follows when more anchor runs narrow the interval.","price":{"kind":"estimate","usd_per_1000_decisions":0.019572722720911453,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 567,051 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0195727227209115/1,000; old list-price estimate USD 0.159947226326394/1,000 (qwen/qwen3.8-27b at USD 0.42/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0195727227209115/1,000. Estimated, not a bill. USD 0.0195727227 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0330328531 over all 1,500 items)."},"latency":{"p50_s_raw":0.08644557849993362,"p50_s_adjusted":0.32289115699986726,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.08644557849993362,"latency_basis":"api-end-to-end","p95_s_raw":0.16484875690017586,"p95_s_adjusted":0.47969751380035175,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.1","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"winnow-12b","name":"Winnow-12B Q8","board":"open","source_url":"https://huggingface.co/EldanRing/Winnow-12B","axes":{"intelligence":59.53366229166759,"calibration":82.95806123333315,"speed":86.16642825724779,"cost":66.95597594555505},"capability":{"score":71.24586176250037,"rank":91,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":68,"api_board_rank":null},"composite_score":72.21241410032238,"rank":71,"open_board_rank":54,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.012631864792722661,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 911,483 input + 0 output tokens over 1,624 decisions = 561.258004926 tokens/decision. GPU-time cost USD 0.0126318647927227/1,000; old list-price cost USD 0.0280629002463054/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0126318647927227/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.1236670445650816,"p50_s_adjusted":0.3973340891301632,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.1236670445650816,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.2292085050139576,"p95_s_adjusted":0.6084170100279152,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":68.88542163459404,"capability":71.24586176250037,"axes":{"calibration":82.95806123333315,"cost":56.556023424877196,"intelligence":59.53366229166759,"speed":86.69753831262085},"rank":30,"open_board_rank":13,"api_board_rank":null}},{"key":"wity-1","name":"wity-1 (Wity, reasoning auto)","board":"api","source_url":"https://wity.alphanimble.com/","axes":{"intelligence":67.88616994428031,"calibration":85.93671111258612,"speed":68.32012851304185,"cost":58.8349720563948},"capability":{"score":76.91144052843322,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.0× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":68.9647088778163,"rank":105,"open_board_rank":null,"api_board_rank":24,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"tariff","usd_per_1000_decisions":0.023559582938388626,"basis":"Wity's stated API tariff USD 0.042 per 1M input tokens, output free; measured on the v1.6.1 common cost basis (all items except those outside the Jev reference input range) from this run's token usage"},"latency":{"p50_s_raw":1.8473002910614014,"p50_s_adjusted":1.8473002910614014,"adjustment":"none (API)","p50_s_measured":1.8473002910614014,"latency_basis":"api-end-to-end","p95_s_raw":7.969840868189922,"p95_s_adjusted":7.969840868189922,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":36,"n_over_limit_answered_ok":36,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.76534010331063,"capability":79.11559492280433,"axes":{"calibration":87.9497550095679,"cost":58.8349720563948,"intelligence":70.28143483604077,"speed":71.79000307789984},"rank":22,"open_board_rank":null,"api_board_rank":14}},{"key":"wity-1-always","name":"wity-1 (Wity, reasoning always)","board":"api","source_url":"https://wity.alphanimble.com/","axes":{"intelligence":67.11413626626955,"calibration":87.14296044619107,"speed":68.5105981536613,"cost":58.8349720563948},"capability":{"score":77.12854835623031,"rank":null,"eligible":false,"within_caps":false,"reasons":["latency 3.1× Jev (v1.5 reference)"],"open_board_rank":null,"api_board_rank":null},"composite_score":69.00315679345155,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":false,"listing":"listed","not_ranked_because":"registry ranked=false or no composite","price":{"kind":"tariff","usd_per_1000_decisions":0.023559582938388626,"basis":"Wity's stated API tariff USD 0.042 per 1M input tokens, output free; measured on the v1.6.1 common cost basis (all items except those outside the Jev reference input range) from this run's token usage"},"latency":{"p50_s_raw":1.9074126482009888,"p50_s_adjusted":1.9074126482009888,"adjustment":"none (API)","p50_s_measured":1.9074126482009888,"latency_basis":"api-end-to-end","p95_s_raw":7.3874665338545995,"p95_s_adjusted":7.3874665338545995,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":36,"n_over_limit_answered_ok":36,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":70.56449594359258,"capability":79.25057518952474,"axes":{"calibration":89.25002486738025,"cost":58.8349720563948,"intelligence":69.25112551166924,"speed":71.20296182763504},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"wity-1-off","name":"wity-1 (Wity, reasoning off)","board":"api","source_url":"https://wity.alphanimble.com/","axes":{"intelligence":44.11116118075945,"calibration":81.82569250484788,"speed":87.61214901323027,"cost":58.8349720563948},"capability":{"score":62.968426842803666,"rank":null,"eligible":false,"within_caps":true,"reasons":[],"open_board_rank":null,"api_board_rank":null},"composite_score":49.18145184761536,"rank":null,"open_board_rank":null,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":false,"listing":"listed","not_ranked_because":"registry ranked=false or no composite","price":{"kind":"tariff","usd_per_1000_decisions":0.023559582938388626,"basis":"Wity's stated API tariff USD 0.042 per 1M input tokens, output free; measured on the v1.6.1 common cost basis (all items except those outside the Jev reference input range) from this run's token usage"},"latency":{"p50_s_raw":0.3454003408551216,"p50_s_adjusted":0.3454003408551216,"adjustment":"none (API)","p50_s_measured":0.3454003408551216,"latency_basis":"api-end-to-end","p95_s_raw":0.501721066609025,"p95_s_adjusted":0.501721066609025,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":49.37499462500363,"capability":62.968426842803666,"axes":{"calibration":81.82569250484788,"cost":58.8349720563948,"intelligence":44.11116118075945,"speed":89.55912090143326},"rank":null,"open_board_rank":null,"api_board_rank":null}},{"key":"xevn-4b","name":"Xevn-4B","board":"open","source_url":"https://huggingface.co/quevirus/Xevn-4B","axes":{"intelligence":52.14523729900746,"calibration":85.13454850231766,"speed":91.64762937997935,"cost":80.50803800213268},"capability":{"score":68.63989290066256,"rank":117,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":90,"api_board_rank":null},"composite_score":73.72484160351587,"rank":51,"open_board_rank":37,"api_board_rank":null,"measured_in":"v1.7.0","draw":{"id":"v1.6-regular-20261012","label":"Second regular draw","release":"v1.7.0","measured_on":"2026-10-24","status":"anchor-runs-pending"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.004464081004051206,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 674,634 input + 0 output tokens over 1,485 decisions = 454.298989899 tokens/decision. GPU-time cost USD 0.00446408100405121/1,000; old list-price cost USD 0.0136289696969697/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00446408100405121/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.04109370599996964,"p50_s_adjusted":0.23218741199993928,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.04109370599996964,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07235619390003763,"p95_s_adjusted":0.29471238780007525,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null},{"key":"xor-26b-a4b","name":"Xor 26B-A4B (Juspay, Gemma-4-26B-A4B, bf16)","board":"open","source_url":"https://huggingface.co/juspay/Xor-26B-A4B","axes":{"intelligence":57.47797916441928,"calibration":84.8928101186779,"speed":90.84255575155115,"cost":72.13803206119073},"capability":{"score":71.18539464154858,"rank":93,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":70,"api_board_rank":null},"composite_score":74.00842889463031,"rank":42,"open_board_rank":29,"api_board_rank":null,"measured_in":"v1.6.1","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.008486571306361034,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: moe gemma, 4B active parameters; throughput 34662.5 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 717,195 input + 1,477 output tokens over 1,477 decisions = 486.57549086 tokens/decision. GPU-time cost USD 0.00848657130636103/1,000; old list-price cost USD 0.0440017941773866/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00848657130636103/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.06005585426464677,"p50_s_adjusted":0.27011170852929356,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.06005585426464677,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.07746530231088398,"p95_s_adjusted":0.30493060462176796,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":23,"n_over_limit_answered_ok":23,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":67.3818936950357,"capability":72.40651985976965,"axes":{"calibration":86.13921620236111,"cost":50.69588844260597,"intelligence":58.673823517178164,"speed":91.03179229063062},"rank":42,"open_board_rank":22,"api_board_rank":null}},{"key":"zerank-2","name":"ZeroEntropy zerank-2","board":"open","source_url":"https://huggingface.co/zeroentropy/zerank-2-reranker","axes":{"intelligence":15.596611891228248,"calibration":83.20040692874471,"speed":81.79510880020156,"cost":60.56419192241651},"capability":{"score":49.39850940998648,"rank":220,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":184,"api_board_rank":null},"composite_score":3.711230315362979,"rank":263,"open_board_rank":214,"api_board_rank":null,"measured_in":"v1.6.0","draw":null,"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.020631316234009007,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 3,409,751 input + 0 output tokens over 1,624 decisions = 2099.60036946 tokens/decision. GPU-time cost USD 0.020631316234009/1,000; old list-price cost USD 0.0524900092364532/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.020631316234009/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases."},"latency":{"p50_s_raw":0.18466085940599442,"p50_s_adjusted":0.5193217188119889,"adjustment":"x2 + 0.15 s (assumption, not measured)","p50_s_measured":0.18466085940599442,"latency_basis":"measured-loopback","p50_s_measured_condition":"on-GPU, no network","latency_note":"Self-hosted models are timed on our own hardware over loopback, without a network, so their measured p50 is not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and the latency cap use the same estimate for every self-hosted row: 2 × measured p50 + 150 ms.","p95_s_raw":0.5618287160526961,"p95_s_adjusted":1.2736574321053922,"n":1500},"method":"v1.7","rescored_in":"v1.7.0","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":0,"n_over_limit_answered_ok":0,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":{"board_release":"v1.6.43","method":"v1.6","composite_score":3.349795360318651,"capability":49.39850940998648,"axes":{"calibration":83.20040692874471,"cost":48.39770052432463,"intelligence":15.596611891228248,"speed":82.89783918855927},"rank":190,"open_board_rank":150,"api_board_rank":null}},{"key":"zils-ai-z3-mini-soup-preview","name":"z3-mini-soup-preview (Zils AI)","board":"open","source_url":"https://huggingface.co/zils-ai/z3-mini-soup-preview","axes":{"intelligence":57.493651530493636,"calibration":86.68299353911392,"speed":93.33840574950976,"cost":82.80646845066562},"capability":{"score":72.08832253480378,"rank":84,"eligible":true,"within_caps":true,"reasons":[],"open_board_rank":61,"api_board_rank":null},"composite_score":77.34016698271256,"rank":13,"open_board_rank":8,"api_board_rank":null,"measured_in":"v1.8.3","draw":{"id":"v1.6-regular-20261027","label":"Third regular draw","release":"v1.8.3","measured_on":"2026-10-31","status":"equated"},"ranked":true,"listing":"ranked","not_ranked_because":null,"price":{"kind":"estimate","usd_per_1000_decisions":0.003742120070964346,"basis":"GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 567,051 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00374212007096435/1,000; old list-price estimate USD 0.0114248018804567/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00374212007096435/1,000. Estimated, not a bill. Reference deepinfra:Qwen/Qwen3.5-4B is frozen at the 25 Sep cut-off and was deprecated afterwards (replaced by Qwen/Qwen3.5-9B); re-score if it changes (I-3). USD 0.00374212007 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.0063155701 over all 1,500 items)."},"latency":{"p50_s_raw":0.026011169000014434,"p50_s_adjusted":0.20202233800002886,"adjustment":"Self-hosted latency estimate under the shared method (assumption, not measured)","p50_s_measured":0.026011169000014434,"latency_basis":"api-end-to-end","p95_s_raw":0.03974401350008065,"p95_s_adjusted":0.2294880270001613,"n":1500},"method":"v1.7","rescored_in":null,"new_in":"v1.8.3","method_status":"v1.7","decision_limit":{"limit_s":10,"n_over_limit":5,"n_over_limit_answered_ok":5,"rule":"latency_s > limit -> failed","speed_basis":"all decisions given (lane items in the raw), measured latency_s","speed_n":1500,"treatment":"share"},"previous":null}]}
