# JevBench by Benchmark Heaven

JevBench measures Jev-class decision models: state and a bounded rubric in, a typed answer out.
Data: JevBench v1.6.39, published 2026-10-22. Release: https://benchmarkheaven.com/jev-models/v1.6.39

## Method
Capability = mean of Intelligence and Calibration, with official ranks only within the cost (0.06459465517241379 USD/1,000 decisions) and median-latency (1.2329566404223442 seconds) caps. Reference: Jev 1.13.0. Composite is secondary.
Latency: API rows are timed end to end, network included. Self-hosted rows are timed on our own hardware over loopback; the ranking and the median-latency cap use an estimate that adds the network round trip (2 x measured p50 + 150 ms), and the latency column shows measured value and estimate. Model authors can ask us to time a hosted endpoint instead (on request).
Axes are benchmark measures, not guarantees. Prices are measured, estimated or self-reported announced values; retain the basis kind. Wrappers and subsidised listings never enter the ranking. Missing values are not published.

## Open-weights board (/jev-models)
Systems we ran ourselves from published weights, ranked by Capability. 140 systems qualify within the caps (Jev 1.13.0 is the unranked reference). Rows outside the caps have no rank.
| Capability rank | Name | Key | Capability | Intelligence | Calibration | Speed | Cost axis | USD per 1,000 decisions | Basis kind | Basis | p50 latency (s) | Source URL |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| reference, not ranked | Jev 1.13.0 (TypeSafe AI) | jev-1.13.0 | 77.12647993588574 | 63.62075784024779 | 90.63220203152369 | 91.49989299046987 | 54.733842199536014 | 0.03227542180094787 | tariff | 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 | 0.23910215869545937 | https://docs.typesafe.ai |
| 1 | Chon-1.0-12B | chon-1.0-12b | 85.33585692323442 | 79.86402006880415 | 90.80769377766468 | 65.53984816070655 | 48.98575014908867 | 0.05017356902356902 | estimate | ESTIMATE: base-model reference price per v1.5-M2 section 2 with INTERPRETATION-1 I-3 (frozen 25 Sep 2026 cut-off snapshots). The exact base google/gemma-4-12B-it (adapter_config base_model_name_or_path, chon_config base_model; pinned commit 707f0a3b) has no listing on OpenRouter or DeepInfra in those snapshots (OPENROUTER-MODELS-20260925.json SHA-256 f5d0cd8e0f7d2d6514748c99bfcaeacc09775bac7aa7eecf765af4ba04b0dcd5; DEEPINFRA-MODELS-20260925.json SHA-256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f) nor in the order's OpenRouter snapshot read 2026-10-19T03:50:11Z, so the nearest listed sibling of the same family and size, OpenRouter google/gemma-3-12b-it, is used: USD 0.05/M input, USD 0.15/M output (DeepInfra lists the same). Same basis as the Cygnet, Winnow-12B, Jev-Omni, Torchcast-Decision-12B and Metask-12B rows. Applied to the server's own usage: input tokens encoded and output tokens generated by its think-when-unsure pass (up to 512 per decision). Self-hosted, no bookable tariff; estimated, not charged. COMMON COST BASIS: published cost is recomputed over the same remaining 1,485 items as every other row; the 15 long states are answered and billed too. All 1,500 items count for Intelligence. | 0.134 measured on-GPU, no network · 0.42 est. incl. network (2x + 150 ms) | https://huggingface.co/ninhnv/chon-1.0-12b |
| 2 | H2O-Lightning-31B v1.0.7 (H2O.ai, Gemma-4-31B-it fine-tune, vLLM + h2oai overlay, one-engine FP8 profile, self-hosted) | h2o-lightning-31b | 85.16947336412102 | 80.15012041049178 | 90.18882631775024 | 89.8424830803211 | 50.57266115000792 | 0.044419939393939394 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 a8a05890ddab2d39d1a1bd543f2b5c83842bdc70aa54766788bc89b7a8a448d7): declared base google/gemma-4-31B-it -> reference() -> google/gemma-4-31b-it, USD 0.09 per million input tokens, USD 0.34 per million output tokens (frozen price cut-off snapshot), the same reference as the other Gemma-4-31B-it rows of this board. The customer asked for the cost of the base model. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the author's server usage block (the full record, 1 output token per decision for the readout). Over all 1,500 items the frozen resolve() gives USD 0.115415 per 1,000 decisions; the official cost axis uses the common 1,485-item basis (below), where the row is under the official cost cap. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 727,319 input / 1,485 output tokens. All-item usage 1,917,924 input / 1,500 output over 1,500 items; the 15 longest items, which this row answered in full, count for no row's cost. | 0.063 measured on-GPU, no network · 0.28 est. incl. network (2x + 150 ms) | https://huggingface.co/h2oai/h2o-lightning-31b |
| 3 | basal-omni-1.0-ultra (Remek Kinas, Gemma-4-31B-it + merged LoRA, typed-decision readout, self-hosted) | basal-omni-1.0-ultra | 82.53799311737751 | 74.05810874808022 | 91.0178774866748 | 87.31015625661291 | 50.37412598110784 | 0.045102 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base google/gemma-4-31B-it -> google/gemma-4-31b-it USD 0.09/M input, USD 0.33999999999999997/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. USD per 1,000 decisions over the common 1,485-item cost basis. Tokens as reported by the author's own server (usage.input_tokens = tokens of the rendered chat prompt incl. system prompt, options and the answer prefill; output_tokens 0, nothing is generated - the model reads the option-code logits at one answer position in one forward pass). The offline count through the author's own renderer and the pinned tokenizer predicted the server's numbers exactly, item by item: 744,183 input tokens over the 1,485 answered items. 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 0.045102 per 1,000 decisions (0.698x the official cost cap USD 0.064595), against USD 0.044651 over all 1,500 items (v1.6.29). | 0.102 measured on-GPU, no network · 0.35 est. incl. network (2x + 150 ms) | https://huggingface.co/Remek/basal-omni-1.0-ultra |
| 4 | Quyet-1.0-Large (Chinh Nguyen, Gemma-4-31B decoder, option-letter logits) | quyet-1-0-large | 81.68880454670384 | 73.40135940607429 | 89.97624968733336 | 86.87440406834797 | 50.546097335562465 | 0.04451059729064039 | estimate | 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): OpenRouter google/gemma-4-31b-it list price $0.09/M input, $0.34/M output (the exact base model google/gemma-4-31B-it; listed since Apr 2026), read 2026-10-04 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 | 0.116 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/chinhnc/Quyet-1.0-Large |
| 5 | Jeff-1.0-Large | jeff_1_0_large | 81.50321747405513 | 71.89886397161271 | 91.10757097649757 | 88.82411852274376 | 49.20157807658991 | 0.04934926977687626 | estimate | Public exact base-model OpenRouter reference for google/gemma-4-31b-it, 2026-10-09: cheapest paid DeepInfra Turbo $0.09/M input, $0.34/M output. Jeff uses native bf16 single forward-pass decision readout with zero generated output tokens. Reference tariff is not an actual Jeff hosting bill. Primary source https://openrouter.ai/api/v1/models/google/gemma-4-31b-it/endpoints; source receipt retained. | 0.080 measured on-GPU, no network · 0.31 est. incl. network (2x + 150 ms) | https://huggingface.co/jgeuter/Jeff-1.0-Large |
| 6 | metask-jev-rain-31B-trained service v3 | metask_jev_rain_31b_trained_v3 | 79.59723052502656 | 72.59243765084337 | 86.60202339920977 | 84.66561996644248 | 49.27770156907625 | 0.04906177777777778 | estimate | v1.5-M2 base-model reference: exact base google/gemma-4-31B-it, OpenRouter google/gemma-4-31b-it list price USD 0.09/M input, USD 0.34/M output (the same reference as the published metask_jev_rain_31b_trained row; snapshot read 2026-10-13T12:10:25Z by the order's public-docs receipt) x the service's own measured usage tokens (840,049 in / 17,397 out); 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 0.049062 per 1,000 decisions (0.760x the official cost cap USD 0.064595), against USD 0.054346 over all 1,500 items (v1.6.29). | 0.181 measured on-GPU, no network · 0.51 est. incl. network (2x + 150 ms) | https://huggingface.co/MeTask-Rain/metask-jev-rain-31B-trained |
| 7 | decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted) | decisio-gemma-4-31b-v080 | 79.56974174390604 | 70.42321091828492 | 88.71627256952715 | 91.18580810090454 | 51.66530928015878 | 0.04084664184157075 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff) and CLEAN: unlike run 54's FP8 sibling, the frozen pricing snapshot resolves this EXACT base - google/gemma-4-31B-it - to USD 0.09 per 1M input and USD 0.34 per 1M output. Read before the pod and asserted in make_registry_r55.py. x this row's own measured tokens, the deck31b method. Nothing is provisional and no release-lane line is open on it. Estimated, not charged | 0.047 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://github.com/aminry/decisio |
| 8 | metask-jev-rain-31B-trained | metask_jev_rain_31b_trained | 78.90688322585615 | 70.71905450578407 | 87.09471194592824 | 83.17884674032656 | 50.73365709012802 | 0.04387442424242424 | estimate | v1.5-M2 base-model reference: exact base google/gemma-4-31B-it (served as Google's unchanged google/gemma-4-31B-it-qat-w4a16-ct@52f3f65b + own LoRA). OpenRouter google/gemma-4-31b-it list price USD 0.09/M input, USD 0.34/M output (listed since Apr 2026; same reference as the published Jeff, deck31b and Quyet-1.0-Large rows), read 2026-10-12T00:13:38Z from https://openrouter.ai/api/v1/models x the service's own measured usage tokens; 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 0.043874 per 1,000 decisions (0.679x the official cost cap USD 0.064595), against USD 0.049211 over all 1,500 items (v1.6.29). | 0.209 measured on-GPU, no network · 0.57 est. incl. network (2x + 150 ms) | https://huggingface.co/MeTask-Rain/metask-jev-rain-31B-trained |
| 9 | metask-jev-rain-12B-trained | metask_jev_rain_12b_trained | 77.87178459131476 | 67.8255189476327 | 87.91805023499681 | 73.37579625429389 | 52.66805268932843 | 0.03782087542087542 | estimate | v1.5-M2 base-model reference: base google/gemma-4-12B-it (author int4 AWQ of google/gemma-4-12B-it-qat-q4_0-unquantized + own LoRA). google/gemma-4-12b-it is not listed on OpenRouter; nearest listed sibling of the same family and size class (fixed Gemma-4-12B precedent) OpenRouter google/gemma-3-12b-it USD 0.05/M input, USD 0.15/M output, read 2026-10-12T00:13:38Z from https://openrouter.ai/api/v1/models x the service's own measured usage tokens; 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 0.037821 per 1,000 decisions (0.586x the official cost cap USD 0.064595), against USD 0.041404 over all 1,500 items (v1.6.29). | 0.115 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/MeTask-Rain/metask-jev-rain-12B-trained |
| 10 | deck-31B (krishna765, frozen Gemma-4-31B-it, TorchAO FP8 dynamic) | deck31b | 77.5907429125261 | 73.02512554551848 | 82.15636027953371 | 86.69970536196755 | 49.69190284016272 | 0.04752658090724441 | estimate | ESTIMATE (v1.5-M2 base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-31b-it list price $0.09/M input, $0.34/M output (the exact frozen base model google/gemma-4-31B-it; same reference as Quyet-1.0-Large), read 2026-10-04 x the server's own measured usage.input_tokens; letter-mass readout, nothing generated; TorchAO FP8 dynamic on 1x GPU (Lium), deck31b 1.0.0; estimated, not charged | 0.127 measured on-GPU, no network · 0.40 est. incl. network (2x + 150 ms) | https://github.com/krishna-gogineni-765/deck31b |
| 11 | René-1 31B FP8 (salfatigroup, Gemma 4 31B full fine-tune, one-pass option readout) | rene-1-31b-fp8 | 75.9540833183004 | 61.72139577013068 | 90.18677086647014 | 87.11720932683018 | 45.963638345890566 | 0.06327207176709546 | estimate | ESTIMATE (base-model reference, not charged). Unmodified pricing_v15.reference resolves google/gemma-4-31B-it to google/gemma-4-31b-it at USD0.09/M input and USD0.34/M output in frozen snapshot. Measured output tokens are 0 by construction (decision_contract.py and release.json); one forward-pass option readout. Common v1.6.1 cost input range applied unchanged. | 0.110 measured on-GPU, no network · 0.37 est. incl. network (2x + 150 ms) | https://huggingface.co/salfatigroup/rene-1-31b-fp8 |
| 12 | H2O-Lightning-4B v1.1 (H2O.ai, Qwen3.5-4B fine-tune, stock vLLM + open shim) | h2o-lightning-4b | 74.99619636935047 | 60.03839906196252 | 89.95399367673842 | 92.61084474720658 | 64.04698652895101 | 0.015791878078817736 | estimate | ESTIMATE (self-hosted open weights, no public tariff): cost per 1,000 decisions measured for H2O-Lightning-4B v1.0 on the v1.5 pool (854,867 input tokens over 1,624 decisions, 0 output tokens; same base, shim and token path) carried for v1.1, the practice of every Qwen3.5-4B row on this board, at USD 0.03 per 1M input and USD 0.15 per 1M output (reference deepinfra:Qwen/Qwen3.5-4B, frozen 25 Sep 2026 snapshot) = USD 0.015792 per 1,000 decisions. Cross-check: v1.1's own v1.6 input tokens on the common cost basis give USD 0.0153 per 1,000 (within 4 %). Corrected in v1.6.37, no new measurement: the earlier cost of this row used USD 0.04 per 1M input; the frozen reference for Qwen/Qwen3.5-4B is USD 0.03 per 1M input and USD 0.15 per 1M output (DeepInfra, 25 Sep 2026 snapshot). | 0.029 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/h2oai/h2o-lightning-4b |
| 13 | 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) | bobcat-flash-1.2 | 73.4936722737144 | 65.81464576505633 | 81.17269878237248 | 90.781726629608 | 49.68577376542501 | 0.04754894380501015 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-26b-a4b-it in the frozen 25 Sep snapshot, USD 0.09 per 1M input and USD 0.3 per 1M output x this row's own measured tokens. Estimated, not charged. THIS IS THE REFERENCE, THE RATE AND THE LABEL THE PUBLISHED PAGE ALREADY CARRIES FOR THIS EXACT BASE on three rows published within the last day - xor-26b-a4b, surogate-rune-26b-a4b-v3-v16 and gev-26b-decide - so pricing this row any other way would make it inconsistent with the field it is placed in. RELEASE-LANE LINE, recorded and not acted on: pricing_v15.reference() RAISES on the author's exact declared spelling google/gemma-4-26B-A4B-it, because that string is in none of the three decision tables, while the lowercase google/gemma-4-26b-a4b-it IS in BASE_REFERENCES and in the dated snapshot at (0.09, 0.30). That is a CAPITALISATION gap - one notch tighter than run 56's -Base/instruct gap on Qwen3.5-35B-A3B - and the three published rows say in their own basis string that the frozen module has no explicit listing-decision line for this base. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: their build_response returns usage.output_tokens == 0 and the readout takes candidate logits at the first answer position of one forward pass, so this row's cost depends on the input rate alone and any other decision is reversible by arithmetic from the raw. | 0.061 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/sanghwa-na/bobcat-flash-1.2 |
| 14 | Surogate Rune 26B-A4B v3 (v1.6 pool) | surogate-rune-26b-a4b-v3-v16 | 73.46807972140657 | 56.055735926690254 | 90.88042351612289 | 86.95393114544444 | 49.597326937281316 | 0.047872830060934325 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-26b-a4b-it in the frozen 25 Sep snapshot, the M2 reference of the published surogate-rune-26b-a4b-v3 row (the frozen module has no explicit listing-decision line for this base: release-lane line), USD 0.09 per 1M input and USD 0.3 per 1M output x this row's own measured tokens. Estimated, not charged | 0.116 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/surogate/rune-26b-a4b-GGUF |
| 15 | Torchcast Decision 4B (v1.2) | torchcast-decision-4b-v1.2 | 72.6066716121222 | 60.262524104890716 | 84.9508191193537 | 91.12622136196717 | 64.8083695127977 | 0.01489547474747475 | estimate | Self-hosted open weights, no bookable tariff. v1.5-M2 base-model reference rule: base Qwen/Qwen3.5-4B (model card base_model, finetune). OpenRouter snapshot of 2026-10-18 lists no Qwen3.5-4B; binding precedent INTERPRETATION-1 I-2/I-3 uses DeepInfra Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output from the frozen 25 Sep 2026 price cut-off snapshot (read 2026-10-18). No-generation readout: output tokens expected 0. Frozen reference deepinfra:Qwen/Qwen3.5-4B; 982,889 input / 0 output tokens over all 1,500 decisions (USD 0.0196578 per 1,000 over all 1,500 items), inside the official cap. 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 0.014895 per 1,000 decisions (0.231x the official cost cap USD 0.064595), against USD 0.019658 over all 1,500 items (v1.6.29). | 0.062 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/torchcast-ai/torchcast-decision-4b |
| 16 | decider-12b v2 (Mapika) | decider-12b | 72.50047627291733 | 63.01250112161242 | 81.98845142422223 | 90.47583056481872 | 57.755449421401664 | 0.025594796798029562 | estimate | add-requests run 46 (v1.5 basis, unpublished) documented hosted-model estimate; no exact base-model floor applies | 0.054 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/Mapika/decider-12b |
| 17 | Xor 26B-A4B (Juspay, Gemma-4-26B-A4B, bf16) | xor-26b-a4b | 72.40651985976965 | 58.673823517178164 | 86.13921620236111 | 91.03179229063062 | 50.69588844260597 | 0.04400179417738659 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-26b-a4b-it in the frozen 25 Sep snapshot, the M2 reference of the published surogate-rune-26b-a4b-v3 row (the frozen module has no explicit listing-decision line for this base: release-lane line), USD 0.09 per 1M input and USD 0.3 per 1M output x this row's own measured tokens. Estimated, not charged | 0.059 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/juspay/Xor-26B-A4B |
| 18 | decider-12b v1, stock Gemma-4-12B-it (Mapika) | decider-12b-v1 | 72.1959364423342 | 60.62893554403879 | 83.7629373406296 | 90.47654290454523 | 57.755449421401664 | 0.025594796798029562 | estimate | add-requests run 46 (v1.5 basis, unpublished) documented hosted-model estimate; no exact base-model floor applies | 0.054 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/Mapika/decider-12b |
| 19 | torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout) | torchcast-decision-12b | 71.73894611323924 | 60.54320091058725 | 82.93469131589123 | 91.74573681567192 | 56.36034169226667 | 0.02848756157635468 | estimate | 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 | 0.037 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://huggingface.co/torchcast-ai/torchcast-decision-12b |
| 20 | Jev-Omni (akhilaaa3, Gemma-4-12B merged) | jev-omni | 71.2873662785918 | 55.530225828826175 | 87.04450672835743 | 85.43173201875288 | 56.05113155939739 | 0.02917173645320197 | estimate | 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 | 0.155 measured on-GPU, no network · 0.46 est. incl. network (2x + 150 ms) | https://huggingface.co/akhilaaa3/Jev-Omni |
| 21 | Winnow-12B Q8 | winnow-12b | 71.24586176250037 | 59.53366229166759 | 82.95806123333315 | 86.69753831262085 | 56.556023424877196 | 0.028062900246305422 | estimate | 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 | 0.117 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/EldanRing/Winnow-12B |
| 22 | Cygnet (blockbrain, frozen Gemma-4-12B-it) | cygnet | 70.90341055515565 | 54.83928122751086 | 86.96753988280044 | 91.77283664982562 | 56.42912569481113 | 0.028337561576354687 | estimate | 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 | 0.037 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://github.com/blockbrain-ai/cygnet-recipe |
| 23 | decisio v0.8.0 on gemma-4-12B-it (frozen, one forward pass, self-hosted) | decisio-gemma-4-12b-v080 | 70.74146614369144 | 52.324988528776075 | 89.15794375860679 | 86.9860428376769 | 59.34583145470655 | 0.022653689911983754 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): the exact base google/gemma-4-12B-it is in the frozen pricing module's UNLISTED_BASE_MODELS set, so no exact base-model rate and no floor exist. The reference is the one the rows PUBLISHED on this same base use - cygnet, jev-omni and winnow-12b - OpenRouter google/gemma-3-12b-it, the nearest publicly hosted 12B Gemma sibling, at its FULL catalog pair USD 0.05 per 1M input and USD 0.15 per 1M output. PREREG-R54-AMENDMENT-1 uses the full pair rather than run 53's output=0, so a row that emits tokens pays for them; where a row generates nothing the output term is zero anyway. x this row's own measured input/output tokens, the deck31b method. Estimated, not charged | 0.102 measured on-GPU, no network · 0.35 est. incl. network (2x + 150 ms) | https://github.com/aminry/decisio |
| 24 | decider chat on Gemma-4-31B-it (Mapika, frozen base, inference technique) | decider-chat-gemma4-31b | 70.58267678405554 | 58.54290392078144 | 82.62244964732962 | 85.78163968405963 | 50.78502198104071 | 0.04370179417738659 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-31b-it, the reference of the published deck31b and quyet-1-0-large rows, USD 0.09 per 1M input and USD 0.34 per 1M output x this row's own measured tokens. Estimated, not charged | 0.126 measured on-GPU, no network · 0.40 est. incl. network (2x + 150 ms) | https://github.com/Mapika/decider |
| 25 | GEV-26B-Decide (AutoTrust, Gemma-4-26B-A4B + LoRA + head) | gev-26b-decide | 70.07613110259072 | 49.817428608862244 | 90.33483359631919 | 89.87310530072351 | 51.12030271996105 | 0.04259153012863913 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter google/gemma-4-26b-a4b-it in the frozen 25 Sep snapshot, the M2 reference of the published surogate-rune-26b-a4b-v3 row (the frozen module has no explicit listing-decision line for this base: release-lane line), USD 0.09 per 1M input and USD 0.3 per 1M output x this row's own measured tokens. Estimated, not charged | 0.052 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/autotrust/GEV-26B-Decide |
| 26 | decisio v0.9.0 on gemma-4-12B-it (frozen, one prefill per question, self-hosted) | decisio-gemma-4-12b-v090 | 69.95438022832886 | 52.299106200484864 | 87.60965425617285 | 85.86778537116794 | 59.34583145470655 | 0.022653689911983754 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff), CARRIED AND NOT NEW: the exact base google/gemma-4-12B-it is in the frozen pricing module's UNLISTED_BASE_MODELS set, so no exact base-model rate and no floor exist, and (0.05, 0.15) is the reference the rows PUBLISHED on this same base already use - cygnet, jev-omni, winnow-12b and torchcast-decision-12b - OpenRouter google/gemma-3-12b-it, the nearest publicly hosted 12B Gemma sibling, at its FULL catalog pair USD 0.05 per 1M input and USD 0.15 per 1M output. It is also the rate THIS ROW'S OWN v0.8.0 SIBLING carries, so the two versions are priced identically and the comparison between them is clean. PREREG-R54-AMENDMENT-1's full pair rather than output=0, x this row's own measured input/output tokens, the deck31b method. The server reports its own usage.input_tokens/output_tokens on every answer (664,759 in and 1,477 out over the answered rows: one prefill, one option-label readout step per decision), so no token-accounting amendment was needed. Estimated, not charged. | 0.132 measured on-GPU, no network · 0.41 est. incl. network (2x + 150 ms) | https://github.com/aminry/decisio |
| 27 | Blink v0.3 26B-A4B | blink-v0-3-26b-a4b-nvfp4 | 69.9453997799703 | 49.21581563112197 | 90.67498392881862 | 91.48151024194223 | 49.62845783533821 | 0.04775858 | estimate | ESTIMATE (*est.), base-model reference. v1.5-M2 base-model reference: Blink v0.3 26B-A4B is a full fine-tune of google/gemma-4-26B-A4B-it quantised to NVFP4 (ModelOpt); the base resolves in pricing_v15 as base_model google/gemma-4-26b-a4b-it -> OpenRouter google/gemma-4-26b-a4b-it, USD 0.09/M input, USD 0.30/M output, from the OpenRouter catalog snapshot 2026-09-25 (jevbench-v15-measure-20260925/receipts/OPENROUTER-MODELS-20260925.json, sha256 f5d0cd8e0f7d2d6514748c99bfcaeacc09775bac7aa7eecf765af4ba04b0dcd5), read by the frozen pricing_v15 on 2026-10-10. Tokens: input = vLLM prompt_tokens, output = 1 per answered question (one decode step, the option letter); both priced. The 29 context-refused decisions carry no tokens and stay in the 1,500-decision denominator. | 0.051 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/PixilabAI/Blink-v0.3-26B-A4B-NVFP4 |
| 28 | Wald 4B v2.1 | wald-4b-v2-1 | 69.27093170132007 | 51.951877017848766 | 86.58998638479137 | 93.07746270925374 | 46.287007162302075 | 0.06172102 | estimate | Self-hosted open weights, no bookable tariff. v1.5-M2 base-model reference rule: base model Qwen/Qwen3.5-4B (model card base_model) -> DeepInfra Qwen/Qwen3.5-4B, USD 0.03/M input and USD 0.15/M output, frozen 25 Sep 2026 price cut-off snapshot (METHOD-v1.5-ADDENDUM-PRICING-INTERPRETATION-1 I-2/I-3; listing deprecated after the cut-off, stays the frozen v1.5 reference). Same basis as the published Wald 4B (v2.0) row. Read 2026-10-10. | 0.027 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/org2ai/Wald-4B |
| 29 | Torchcast Decision 4B (v1.1) | torchcast-decision-4b-v1.1 | 68.66260515667253 | 54.20841827353793 | 83.11679203980711 | 93.62765623388151 | 64.8083695127977 | 0.014895474747474746 | estimate | ESTIMATE, v1.5-M2 base-model reference rule: self-hosted open weights, no bookable tariff. Base model Qwen/Qwen3.5-4B (model card base_model). OpenRouter does not list Qwen3.5-4B (host snapshot source/public-docs/OPENROUTER-MODELS.json, fetched 2026-10-12T08:15:26Z, read 2026-10-12). Binding precedent INTERPRETATION-1 I-2/I-3: DeepInfra Qwen/Qwen3.5-4B, USD 0.03/M input, USD 0.15/M output, frozen 25 Sep 2026 price cut-off snapshot (listing later deprecated; frozen reference kept). Logit readout, no generation: server reports 0 output tokens. 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 0.014895 per 1,000 decisions (0.231x the official cost cap USD 0.064595), against USD 0.019658 over all 1,500 items (v1.6.29). | 0.021 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/torchcast-ai/torchcast-decision-4b |
| 30 | Hopper 12B trained (gemma-4-12B-it frozen + LoRA r32 unmerged, one forward pass, self-hosted) | hopper-12b-trained | 68.14610127602074 | 51.6576814732076 | 84.63452107883388 | 84.57301814857726 | 56.15749203571245 | 0.02893456330399459 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): the exact base google/gemma-4-12B-it is in the frozen pricing module's UNLISTED_BASE_MODELS set, so no exact base-model rate and no floor exist. The reference is the one the rows PUBLISHED on this same base use - cygnet, jev-omni and winnow-12b - OpenRouter google/gemma-3-12b-it, the nearest publicly hosted 12B Gemma sibling, at its FULL catalog pair USD 0.05 per 1M input and USD 0.15 per 1M output. PREREG-R54-AMENDMENT-1 uses the full pair rather than run 53's output=0, so a row that emits tokens pays for them; where a row generates nothing the output term is zero anyway. x this row's own measured input/output tokens, the deck31b method. Estimated, not charged | 0.164 measured on-GPU, no network · 0.48 est. incl. network (2x + 150 ms) | https://huggingface.co/HopitAI/hopper |
| 31 | SPX-CD-Omni (SurdAI, google/gemma-4-12B-it + LoRA r32, one forward pass, self-hosted) | spx-cd-omni | 68.1286624051734 | 48.779308810848946 | 87.47801599949784 | 89.44891259349035 | 58.27072766098771 | 0.024602301963439405 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): the exact base google/gemma-4-12B-it is in the frozen pricing module's UNLISTED_BASE_MODELS set, so no exact base-model rate and no floor exist. The reference is the one the three rows PUBLISHED on this same base all use - cygnet, jev-omni and winnow-12b - OpenRouter google/gemma-3-12b-it list price USD 0.05 per 1M input, USD 0.00 per 1M output (the nearest publicly hosted 12B Gemma sibling; output is 0 because the runner reads candidate logits in one forward pass and generates nothing), x this row's own measured usage.input_tokens, the deck31b method. Estimated, not charged | 0.073 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://huggingface.co/SurdAI/SPX-CD-Omni |
| 32 | Aplomb 1 (5.3B decision model on Qwen3.5-4B, trained readout head, self-hosted) | aplomb-1 | 67.51828279493242 | 45.81854920739271 | 89.21801638247211 | 88.88310989631262 | 57.86508463809988 | 0.025380324983073795 | estimate | ESTIMATE (base-model reference; self-hosted open weights): the exact declared base Qwen/Qwen3.5-4B resolves in the frozen pricing snapshot through DEEPINFRA_BASE_REFERENCES to USD 0.03 per 1M input and USD 0.15 per 1M output, read before the pod and asserted in make_registry_r55.py. The same rate run 54 gave messier-one, the other self-hosted Qwen3.5-4B derivative. DELIBERATELY NOT the vendor's own cheaper public tariff: Aplomb 1 is also sold on EmpirioLabs' hosted API at USD 0.02 per 1M input with output free (GitHub #195), which would LOWER this row's cost - but all 106 self-hosted rows in the registry carry cost_kind `estimate` and not one is priced at a vendor tariff, and creating that precedent for one row is the release lane's call, not this job's. The estimate can therefore only OVERSTATE this row, and it is reversible by arithmetic because the raw keeps the tokens. x this row's own measured tokens. INPUT TOKENS ARE EXACT, NOT ESTIMATED, and they include BOTH passes of the author's own --debias flag on every multi-option choice, because that is what the configuration consumes. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the author's answer_logits runs the model with logits_to_keep=1, use_cache=False and reads label logits, and nothing is generated anywhere in their code path. Counts carried in the scorer's own cost_estimate field, never written into usage. Estimated, not charged | 0.094 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://huggingface.co/empiriolabsai/aplomb-1 |
| 33 | Bespoke Nimble 9B v3 (Qwen3.5-9B LoRA) | bespoke-nimble-9b-v3 | 66.79290880490993 | 56.54011796945148 | 77.04569964036838 | 89.19502471208428 | 46.094951684283686 | 0.06263757616790792 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter qwen/qwen3.5-9b, the reference of the published nimble-9b row (same base), USD 0.1 per 1M input and USD 0.15 per 1M output x this row's own measured tokens. Estimated, not charged | 0.094 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://huggingface.co/bespokelabs/Bespoke-Nimble-9B-v3 |
| 34 | Metask rain 4B | metask-jev-rain-4b | 65.9864012915813 | 51.08240263971605 | 80.89039994344657 | 79.59790373470649 | 51.82161587751452 | 0.04035953346855983 | estimate | Self-hosted open weights, no bookable tariff. v1.5-M2 base-model reference: the model card names base_model Qwen/Qwen3.5-4B. Per the frozen v1.5 precedent (METHOD-v1.5-ADDENDUM-PRICING-INTERPRETATION-1.md I-2/I-3), Qwen3.5-4B rows use DeepInfra Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output (frozen 25 Sep 2026 price cut-off snapshot; listing later deprecated, frozen reference retained per I-3). Read from the official method files on 2026-10-08; no fresh provider page was fetched. Output tokens priced at the reference output rate. Estimate, not a bookable tariff. | 0.058 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/wayfind/metask-jev-4b-policy-mix |
| 35 | shisa-de-1 (shisa-ai, Gemma-4-26B-A4B) | shisa-de-1 | 65.54874380543649 | 49.89769839508435 | 81.19978921578861 | 92.81832911721798 | 49.76802917325137 | 0.04724969696969697 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base google/gemma-4-26b-a4b-it -> google/gemma-4-26b-a4b-it USD 0.09/M input, USD 0.3/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the evaluator's adapter (prompt tokens of the request(s); output_tokens 0 - the single sampled token is discarded, the answer is the logprob readout). Base string is the frozen module's lower-case key (the card spelling raises). 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 0.047250 per 1,000 decisions (0.731x the official cost cap USD 0.064595), against USD 0.046777 over all 1,500 items (v1.6.29). | 0.031 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/shisa-ai/shisa-de-1 |
| 36 | Plumb-4B (crh225, JevK5 v0.2 + LoRA) | plumb-4b | 65.42184422831923 | 42.9626698747687 | 87.88101858186977 | 93.83356539601934 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.020 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/crh225/plumb-4b |
| 37 | Arcbright Elf 1 12B | arcbright-elf-1-12b | 64.96139735769383 | 43.75476217706713 | 86.16803253832052 | 91.97994026431576 | 57.38209314881071 | 0.026338855218855223 | estimate | ESTIMATE (*est.), nearest-listed-sibling reference: declared base google/gemma-4-12B-it has no market listing (UNLISTED in the frozen pricing_v15 module, sha256 6202ef78a2c83cbffc9389cf1f37df70fe0c85ed03f4dd5c5f997f3d31aae13a), priced through the frozen catalog's nearest listed sibling of the same family and size class: google/gemma-3-12b-it USD 0.05/M input, USD 0.15/M output (the fixed Gemma-4-12B precedent, same basis as the published metask_jev_rain_12b_trained row). Estimated, not charged, not a quote from the author or any host. USD per 1,000 decisions on the denominator of the release basis (the same-items cost rule for this draw prices the common item set without the 15 very long items the author's server refuses, for every row alike). Tokens as reported by the author's own server usage block (input_tokens = tokens of the rendered chat prompt incl. system prompt, options and the answer prefill; output_tokens = 1, exactly one vLLM forward pass per answered single-question row - nothing is generated, the option-code logits are read in that one pass). The offline count through the author's own render.py and the pinned tokenizer predicted the server's numbers EXACTLY, item by item: 777,809 input tokens over the 1,485 answered items, and 15 HTTP 422 refusals at the pre-registered ids and positions. | 0.034 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://huggingface.co/arcbright/elf-1-12b |
| 38 | decider-4b v2 (Mapika) | decider-4b-v2 | 64.88166453375648 | 40.12597268917964 | 89.63735637833331 | 91.78549329698745 | 64.54376873042901 | 0.015201077586206898 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.025 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://github.com/Mapika/decider |
| 39 | JevK5 v0.3 (4B) | jevk5-v0.3-4b | 64.20716028123837 | 38.53201254374527 | 89.88230801873149 | 93.9472917853559 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.019 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/alibiserikbay/JevK5 |
| 40 | Clef-Flash (Cloudflare, Qwen3.5-9B post-train with a joint schema head, multimodal, measured on text) | clef-flash | 63.824241207952895 | 42.18368888957246 | 85.46479352633332 | 88.4044196893156 | 46.79963593205177 | 0.05933971674876847 | estimate | latest published v1.5 row (v1.5.8 combined file) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied; self-hosted estimate, not measured hosted billing | 0.103 measured on-GPU, no network · 0.36 est. incl. network (2x + 150 ms) | https://huggingface.co/Cloudflare/clef-flash |
| 41 | JPT-4B (kirp, Qwen3.5-4B LoRA) | jpt-4b | 63.820192644345624 | 40.7504160092812 | 86.88996927941005 | 87.95549105905897 | 64.21228687378368 | 0.01559278787878788 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server, KEPT AS SENT: output_tokens is 1 per answered question (the label position the engine counts, not generated text); this adds about USD 0.0002 per 1,000 decisions at the reference output rate. 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 0.015593 per 1,000 decisions (0.241x the official cost cap USD 0.064595), against USD 0.015437 over all 1,500 items (v1.6.29). | 0.063 measured on-GPU, no network · 0.28 est. incl. network (2x + 150 ms) | https://huggingface.co/kirp/jpt-4b |
| 42 | deck-4B v1.0 (krishna765, JevK5 + LoRA, FP8 weight-only) | deck-4b-v1-0 | 63.60791322392738 | 38.62465958342338 | 88.59116686443139 | 89.93848373337222 | 63.46145558281372 | 0.01651777251184834 | estimate | ESTIMATE (v1.5-M2 base-model reference; self-hosted open weights, no public tariff): DeepInfra Qwen/Qwen3.5-4B list price $0.03/M input, $0.15/M output (base of JevK5, the adapter's base checkpoint; not on OpenRouter; frozen 25 Sep snapshot per I-3, same reference as reflex-4b, jobe, typecastlm and Quyet) x the server's own measured usage.input_tokens; logit readout, 0 output tokens; FP8 weight-only on 1x RTX PRO 6000 96GB (Lium; v1.6 GPU-class deviation from source H100), deck-4b 1.0.0 + jevk5 f944fe37; estimated, not charged | 0.076 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://github.com/krishna-gogineni-765/deck |
| 43 | Jebadiah 9B v2.1 (Frontier Infra, Qwen3.5-9B LoRA) | jebadiah-9b-v2-1 | 63.41348930844194 | 37.15024789587638 | 89.67673072100749 | 91.39972504887733 | 48.85607041331758 | 0.05067545454545454 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-9B -> qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 533d80f0...d960e5, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens as reported by the author's own server (750,303 input over the 1,485 answered items, output_tokens 1 per answered request). 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 0.050675 per 1,000 decisions (0.785x the official cost cap USD 0.064595), against USD 0.050169 over all 1,500 items (v1.6.29). | 0.043 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/frontier-infra/jebadiah-9b-v2-1 |
| 44 | Seb-9B (ironbcc, Qwen3.5-9B fine-tune, one forward pass, stock vLLM, self-hosted) | seb-9b | 63.13800598261042 | 38.312942577836566 | 87.96306938738428 | 90.8001817381692 | 48.88742799396291 | 0.050553636363636356 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 6202ef78a2c83cbffc9389cf1f37df70fe0c85ed03f4dd5c5f997f3d31aae13a): base Qwen/Qwen3.5-9B, now declared by the author on the card (revision 4ae564d6) -> qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by vLLM's own usage block: 748,494 input on the 1,485 answered items (equal to our offline per-item count item by item, pre-registered), 1 output token per answered row (one forward pass that reads the label logprobs); the 15 refused items count 0 tokens. | 0.050 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/ironbcc/seb-9b |
| 45 | janus 4B (Icarus AI / cmxu, Qwen3.5-4B + LoRA r64 and pointer decision head) | janus-4b | 62.85256571162924 | 37.57770128683752 | 88.12743013642095 | 93.17471817821269 | 64.41461114142905 | 0.015352518472906404 | estimate | add-requests run 47 (v1.5 basis, unpublished) base-model price floor deepinfra:Qwen/Qwen3.5-4B | 0.030 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/cmxu/janus-4b |
| 46 | JevK5 v0.2.0 | jevk5-v02 | 62.824112319658205 | 36.261123811123895 | 89.38710082819252 | 91.56570599843724 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.040 measured on-GPU, no network · 0.23 est. incl. network (2x + 150 ms) | https://github.com/allebee/jevk5 |
| 47 | pplx-decider-flash-v1.2 (Perplexity, Qwen3.5-4B full fine-tune, marker readout, self-hosted) | pplx-decider-flash-v1.2 | 62.51340446094119 | 46.343042310166375 | 78.68376661171601 | 91.76093007933393 | 64.42558238272366 | 0.015339595959595961 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 6202ef78a2c83cbffc9389cf1f37df70fe0c85ed03f4dd5c5f997f3d31aae13a): declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. USD per 1,000 decisions over the full 1,500-item denominator. Tokens as reported by the author's own `Decider.decide()` usage block (input = every non-padding position the request used; output 0, nothing is generated - the logits are read at the markers in one forward pass). The offline count through the author's own protocol.Encoder predicted the server's numbers EXACTLY, item by item: 1,939,205 input tokens over all 1,500 answered items, and the same prediction held on the 231 public items (154,452) before any sealed item was dispatched. COMMON COST BASIS: the preceding full-1,500-item figure is an all-items diagnostic. The published cost in this release uses the same remaining 1,485 items as every other row, under the unchanged v1.6.30 cost rule; USD 0.015339595959595961 per 1,000 decisions. All items still count for Intelligence. | 0.042 measured on-GPU, no network · 0.23 est. incl. network (2x + 150 ms) | https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.2 |
| 48 | Messier One v0.4 (Qwen3.5-4B fine-tune, self-hosted) | messier-one-v0.4 | 62.443616494634156 | 52.02542301918905 | 72.86180997007926 | 89.67127141858641 | 61.121663950171616 | 0.01976717171717172 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 dcfcecb0..., the module in force; its base-model references are unchanged since the dae86740... module this row was first priced with - the two later re-freezes each added one hosted-API launch-list key and nothing else): declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output, the same basis the live messier-one-v0.2 row carries. The author's server reports NO usage block of any kind, so the input tokens are counted OFFLINE with the author's own Prompter (prompt.py at tag v0.4) at the exact serving settings (--max-state 7500 --max-total 8150, AA labels, qboth layout) and cross-checked against vLLM's own prompt-token counters from the measured pass; output tokens are 0 (the engine is asked for one token restricted to the option labels and nothing is generated for the caller). The reference listing was deprecated after the 25 Sep 2026 cut-off and replaced by Qwen/Qwen3.5-9B; the frozen snapshot rates remain in use (I-3). Estimated, not charged, not a quote from the author or any host. 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 0.019767 per 1,000 decisions (0.306x the official cost cap USD 0.064595), against USD 0.022118 over all 1,500 items (v1.6.29). | 0.066 measured on-GPU, no network · 0.28 est. incl. network (2x + 150 ms) | https://huggingface.co/agentmessier/messier-one |
| 49 | Hopper | hopper | 62.40936838189898 | 34.73970870443837 | 90.07902805935959 | 91.11434937226889 | 62.25627699896333 | 0.018118577586206896 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.056 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/HopitAI/hopper |
| 50 | Imajev-4B (RTX 5090) | imajev-4b-rtx5090-a2 | 61.89494846465412 | 34.62714018309005 | 89.1627567462182 | 90.51878000442085 | 63.26458220826236 | 0.016769261083743846 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (I-2): measured input tokens; zero generated output tokens for signed logits readout. M2 floor uses the 25 Sep 2026 DeepInfra Qwen3.5-4B snapshot rates (USD 0.03/M input, USD 0.15/M output); the snapshot records deprecation on 11 Jun 2026 and replacement by Qwen3.5-9B. | 0.071 measured on-GPU, no network · 0.29 est. incl. network (2x + 150 ms) | https://github.com/fstandhartinger/jevbench/issues/110 |
| 51 | jqv (Qwen3-32B zero-shot) | jqv | 61.8647697220168 | 33.85815624201783 | 89.87138320201578 | 83.1032423977654 | 51.172592607797476 | 0.042420935960591134 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.097 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://github.com/Octalab-Inc/jqv |
| 52 | metask-jev-4b | metask-jev-4b | 61.79110407115073 | 39.166454612935595 | 84.41575352936586 | 90.78239008886905 | 57.7311327584882 | 0.025642610837438424 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.054 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://github.com/metask-ai/metask-jev |
| 53 | Kahn1 4B (Okura66, Qwen3.5-4B LoRA merge, author's sysone engine) | kahn1 | 61.74359439203453 | 39.71971801341685 | 83.76747077065221 | 89.97531755002697 | 50.12545250055582 | 0.04597110221674877 | estimate | ESTIMATE, ours (the author's server returns no token usage; Florian 7 Oct 2026: use our own estimate): 829,523 input tokens counted with the author's tokenizer and prompt builder over the 1,624 v1.5 items x n_permutations 3 (the author's server default) = 2,488,569 tokens / 1,624 decisions x USD 0.03 per 1M input and USD 0.15 per 1M output (deepinfra:Qwen/Qwen3.5-4B, frozen 25 Sep 2026 snapshot), 0 output tokens = USD 0.045971 per 1,000 decisions; carried like every Qwen3.5-4B row on this board. Cross-check: the same count on the v1.6 common cost basis (1,477 items, x3) gives USD 0.0448 per 1,000 (within 3 %). Corrected in v1.6.37, no new measurement: the earlier cost of this row used USD 0.04 per 1M input; the frozen reference for Qwen/Qwen3.5-4B is USD 0.03 per 1M input and USD 0.15 per 1M output (DeepInfra, 25 Sep 2026 snapshot). | 0.066 measured on-GPU, no network · 0.28 est. incl. network (2x + 150 ms) | https://huggingface.co/Okura66/Kahn1-Qwen3.5-4B |
| 54 | lev (Interfaze AI, Qwen3.5-4B + LoRA r32, label-token readout + candidate-path head, per-bucket temperatures) | lev | 61.70420936973613 | 41.084702071234844 | 82.32371666823741 | 88.00251499520694 | 61.79891447724963 | 0.01876590517241379 | estimate | latest published v1.5 row (v1.5.8 combined file) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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); self-hosted estimate, not measured hosted billing | 0.102 measured on-GPU, no network · 0.35 est. incl. network (2x + 150 ms) | https://huggingface.co/interfaze-ai/lev |
| 55 | Decision 2.0 Lux 9B (vLLM-SR, Qwen3.5-9B lineage) | decision2-lux-9b | 61.64951208707873 | 35.82270435669186 | 87.47631981746561 | 89.93254869015362 | 48.45008367155653 | 0.05227939393939394 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: lineage base Qwen/Qwen3.5-9B (the card's base vllm-sr/Decision-1.0-Lux-9B is itself a fine-tune and RAISES in the module; Qwen3.5-9B lineage verified by config comparison in the earlier v1.5 measurement) -> qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens as reported by the author's runtime (output_tokens 0, nothing generated). 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 0.052279 per 1,000 decisions (0.809x the official cost cap USD 0.064595), against USD 0.051757 over all 1,500 items (v1.6.29). | 0.068 measured on-GPU, no network · 0.29 est. incl. network (2x + 150 ms) | https://huggingface.co/vllm-sr/Decision-2.0-Lux-9B |
| 56 | blink-4b (thegovind, Qwen3.5-4B derivative) | blink-4b | 61.62516934780887 | 35.29148713463226 | 87.95885156098548 | 91.30713992573452 | 63.5219850233163 | 0.01644121212121212 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server (output_tokens 0, nothing generated). 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 0.016441 per 1,000 decisions (0.255x the official cost cap USD 0.064595), against USD 0.039886 over all 1,500 items (v1.6.29). | 0.057 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/thegovind/blink-4b |
| 57 | Malkuth-4B (newfull5, Kev post-train) | malkuth-4b | 61.139059860562405 | 33.840703805987175 | 88.43741591513763 | 89.51571180654442 | 55.82433433950372 | 0.02968398399014778 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied; 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) | 0.074 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://github.com/newfull5/malkuth |
| 58 | Decision 4B v1.2 (FlyMyJev, Qwen3.5-4B + LoRA) | decision-4b-v12 | 60.41399746452394 | 32.40310972936693 | 88.42488519968094 | 93.92936299294371 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.019 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://benchmarkheaven.com/jev-models/decision-4b-v12 |
| 59 | jiwo 4b (Qwen3.5-4B backbone with a trained decision head, one forward pass, native probabilities, self-hosted) | jiwo-4b | 60.227051544327715 | 34.784391844693545 | 85.66971124396188 | 92.44547609658223 | 64.0899740452639 | 0.01573986 | estimate | ESTIMATE (*est.), exact base-model market reference; self-hosted open weights with no per-token tariff of their own. The unmodified frozen pricing_v15 resolves the author's own declared base spelling Qwen/Qwen3.5-4B straight through to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the dated 25 Sep 2026 snapshot. No new release-lane pricing line.; reference `deepinfra:Qwen/Qwen3.5-4B` frozen at the 25 Sep cut-off, deprecated afterwards (replaced by `Qwen/Qwen3.5-9B`); re-score if it changes (I-3) OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 786,993 input and 0 output tokens over 1,500 decisions. Estimated, not charged. | 0.034 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://huggingface.co/eljiwo/jiwo-4b |
| 60 | Wald 4B | wald-4b-v2 | 59.75558427184466 | 41.11158964477434 | 78.39957889891498 | 82.20577865115169 | 54.30361484501323 | 0.03335898580121703 | estimate | Self-hosted open weights, no bookable tariff. v1.5-M2 base-model reference rule: base model Qwen/Qwen3.5-4B (model card base_model) -> DeepInfra Qwen/Qwen3.5-4B, USD 0.03/M input and USD 0.15/M output, frozen 25 Sep 2026 price cut-off snapshot (METHOD-v1.5-ADDENDUM-PRICING-INTERPRETATION-1 I-2/I-3; listing deprecated after the cut-off, stays the frozen v1.5 reference). Output tokens priced at the reference output price. Read 2026-10-08. | 0.031 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/org2ai/Wald-4B |
| 61 | Quyet-1.0-Medium (Chinh Nguyen, Qwen3.5-4B decoder, option-letter logits) | quyet-1-0-medium | 59.24679183153445 | 41.73350604436517 | 76.76007761870373 | 89.46638461924755 | 63.43457194636937 | 0.01655189039408867 | estimate | 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 Qwen/Qwen3.5-4B list price $0.03/M input (the exact base weights; not on OpenRouter; same reference as reflex-4b, jobe and typecastlm), read 2026-09-24 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 | 0.077 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://huggingface.co/chinhnc/Quyet-1.0-Medium |
| 62 | Decision-4B (Eval Engine / Chromia) | evalengine-decision-4b | 59.09507049015456 | 37.93489123230218 | 80.25524974800695 | 91.06043780349009 | 65.87776871337948 | 0.013721693349753695 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: documented hosted-model estimate; reference deepinfra:Qwen/Qwen3.5-4B frozen at the 25 Sep cut-off, already marked deprecated at that cutoff (replaced by Qwen/Qwen3.5-9B); re-score if it changes (I-3). Frozen market reference deepinfra:Qwen/Qwen3.5-4B. | 0.047 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/evalengine/decision-4b-gguf |
| 63 | Manchego v2.1 | manchego-v21-v15 | 58.27541943544253 | 32.54303915979197 | 84.0077997110931 | 90.7347510499745 | 64.32996474661243 | 0.015452586206896551 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: frozen 25 Sep DeepInfra Qwen/Qwen3.5-4B market reference, already marked deprecated at that cutoff (replaced by Qwen/Qwen3.5-9B). $0.03/M input, zero generated output; 836,500 measured input tokens across 1,624 decisions. Frozen v1.5 base-model floor applied; no bookable Manchego tariff claimed. Re-score if the basis changes. | 0.054 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://github.com/nschlaepfer/manchego-serve |
| 64 | Messier One v0.2 (Qwen3.5-4B fine-tune, one prefill, self-hosted) | messier-one-v0.2 | 58.130562443744644 | 42.35139759222662 | 73.90972729526267 | 92.70909408602279 | 64.78240136302922 | 0.014925192958700068 | estimated | ESTIMATE (base-model reference; self-hosted open weights). NO JUDGEMENT OF OURS: pricing_v15.reference() resolves Qwen/Qwen3.5-4B straight through to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the frozen 25 Sep snapshot (the same rate aplomb-1 carries for the same base) x this row's own measured tokens. Estimated, not charged. THE FRONT END REPORTS NO TOKEN USAGE (usage {}), exactly as v0.1 did in run 54, so the counts come from run 54's pre-registered route unchanged (PREREG-R59-AMENDMENT-1 = R54-AMENDMENT-2): the author's own prompt.Prompter (v0.2 source, unmodified) recomputes each item's prompt token ids, output is fixed at 1 by their own max_tokens=1, and the totals were required to match vLLM's own counters or the row would get no cost: 914,665 input and 1,500 output tokens both ways (vllm prompt_tokens_total 915,349 - 684; generation_tokens_total 1,508 - 8). Counts sit in cost_estimate, never in usage. | 0.027 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/agentmessier/messier-one |
| 65 | Diffusion Jev (DiffusionGemma 26B-A4B on patched SGLang, 48-step diffusion readout, self-hosted) | diffusion-jev | 58.08750812795111 | 54.860121638521335 | 61.31489461738089 | 86.79193858661813 | 49.559714426383636 | 0.04801123222748815 | estimate | ESTIMATE, and the caveat travels with the number rather than sitting in a footnote. The frozen pricing snapshot returns None for google/diffusiongemma-26B-A4B-it (it is unlisted: no exact rate and no floor), and a scan of all 129 registry entries finds NO published row on any DiffusionGemma base at all - the page's openjev-razorback16 is the same model but a different submitter and carries no v1.6 score - so there was no precedent to copy. The reference used is google/gemma-4-26b-a4b-it, USD 0.09 per 1M input and USD 0.30 per 1M output: the identically sized and identically shaped publicly hosted Gemma 4 26B-A4B sibling of the unlisted diffusion checkpoint, a closer sibling than the one run 54 had to use for gemma-4-12B-it. x this row's own measured tokens. THE CAVEAT: a token-priced cost axis charges this row as though it ran one forward pass, when the author's own documentation denoises a 256-position canvas for up to 48 adaptive steps per question. The Cost axis therefore UNDERSTATES this configuration's compute, by construction and not by accident; the measured per-question denoising_steps are kept in the raw so the understatement is quantifiable. Estimated, not charged | 0.131 measured on-GPU, no network · 0.41 est. incl. network (2x + 150 ms) | https://huggingface.co/google/diffusiongemma-26B-A4B-it |
| 66 | jev-local (Qwen3.5-9B) | jev-local | 57.458381676033426 | 41.85041350559534 | 73.06634984647151 | 75.55955847735143 | 58.85554343293177 | 0.023522413793103446 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.365 measured on-GPU, no network · 0.88 est. incl. network (2x + 150 ms) | https://github.com/us/jev-local |
| 67 | Decision 4B v1.1 (FlyMyJev, Qwen3.5-4B + LoRA) | decision-4b-v11 | 57.18675493386165 | 29.134168280172865 | 85.23934158755043 | 93.95641941218838 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.019 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://benchmarkheaven.com/jev-models/decision-4b-v11 |
| 68 | JEV Qwen3.5-9B Base NVFP4 | jev-qwen3.5-9b-base-nvfp4 | 57.14200691217819 | 29.251131763478327 | 85.03288206087805 | 93.62779491136789 | 47.592235403520306 | 0.05583743842364531 | estimate | 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 | 0.023 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/WIlfLin/JEV-Qwen3.5-9B-Base-NVFP4 |
| 69 | iota-decide-4b (Macrocosmos, Clef joint-head 4B, self-hosted) | iota-decide-4b | 57.12433965292368 | 29.414053145614563 | 84.8346261602328 | 87.4670144658746 | 62.97034212213745 | 0.017152282828282833 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 a8a05890ddab2d39d1a1bd543f2b5c83842bdc70aa54766788bc89b7a8a448d7): base Qwen/Qwen3.5-4B-Base priced as Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the server usage block: 1,094,798 input on 1,500 answered items, 0 output tokens (one forward pass, joint-head readout, no generation). COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 849,038 input / 0 output tokens. All-item usage 1,094,798 input / 0 output over 1,500 items. | 0.117 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/macrocosm-os/iota-decide-4b |
| 70 | Uniphore-4B-Decider | uniphore-4b-decider | 57.0653357969117 | 34.15161183674326 | 79.97905975708014 | 87.70990379280212 | 64.54050247464266 | 0.01520488888888889 | estimate | v1.5-M2 base-model reference: frozen DeepInfra snapshot jevbench-v15-measure-20260925 receipts/DEEPINFRA-MODELS-20260925.json (SHA-256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f); listing Qwen/Qwen3.5-4B (exact base, merge_info.json base_revision 851bf6e8), USD 0.03/M input and USD 0.15/M output. Note (METHOD-v1.5-ADDENDUM-PRICING-INTERPRETATION-1 I-3): the frozen DeepInfra Qwen/Qwen3.5-4B entry is flagged deprecated (replaced_by Qwen/Qwen3.5-9B) while still showing this price; it remains the frozen reference. The OpenRouter /api/v1/models snapshot in source/public-docs (fetched 2026-10-12T07:05:09Z) lists no Qwen3.5-4B route. Self-hosted GPU deployment has no provider tariff, so price_kind is estimate. The native readout is one causal forward pass with no generated tokens: usage.output_tokens is 0 by construction (logits_to_keep=1). 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 0.015205 per 1,000 decisions (0.235x the official cost cap USD 0.064595), against USD 0.015053 over all 1,500 items (v1.6.29). | 0.101 measured on-GPU, no network · 0.35 est. incl. network (2x + 150 ms) | https://huggingface.co/Uniphore/uniphore-4b-decider |
| 71 | Jebadiah 9B v2 (Frontier Infra, Qwen3.5-9B LoRA) | jebadiah-9b-v2 | 56.949321883834315 | 30.5871791271758 | 83.31146464049283 | 91.57148361254983 | 48.85607041331758 | 0.05067545454545454 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-9B -> qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server, KEPT AS SENT: output_tokens is 1 per answered question (the label position the engine counts, not generated text); this adds about USD 0.0002 per 1,000 decisions at the reference output rate. 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 0.050675 per 1,000 decisions (0.785x the official cost cap USD 0.064595), against USD 0.050169 over all 1,500 items (v1.6.29). | 0.054 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/frontier-infra/jebadiah-9b-v2 |
| 72 | CoCo-Decision-4B (corners-ai, LoRA on Qwen3.5-4B, option-label logits of one pass, served by oh-my-jev) | coco-decision-4b | 56.92005372388988 | 32.65682750200105 | 81.18327994577871 | 91.3615355688411 | 65.17140064644872 | 0.014486161137440754 | estimate | ESTIMATE (base-model reference; self-hosted open weights). NO JUDGEMENT OF OURS: pricing_v15.reference() resolves Qwen/Qwen3.5-4B straight through to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the frozen 25 Sep snapshot (the same rate aplomb-1 carries for the same base) x this row's own measured tokens. Output tokens are 0 in every answered record: one prefill, option-label readout, nothing generated. Estimated, not charged. | 0.044 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/corners-ai/CoCo-Decision-4B-Ko |
| 73 | APUS-OpenJev-v1-9B (merged bf16 checkpoint-3000 on Qwen3.5-9B, their own native runtime, full depth, self-hosted) | apus-openjev-v1-9b | 56.907845099085506 | 47.20514429531085 | 66.61054590286017 | 82.18245466956745 | 48.738123569705756 | 0.05113628977657414 | estimate | ESTIMATE (base-model reference; self-hosted open weights): the exact declared base Qwen/Qwen3.5-9B resolves in the frozen pricing snapshot through BASE_REFERENCES to qwen/qwen3.5-9b = USD 0.10 per 1M input and USD 0.15 per 1M output, read before the pod and asserted in make_registry_r56.py. Nothing provisional. x this row's own measured tokens. BECAUSE OUTPUT TOKENS ARE 0 BY CONSTRUCTION this axis rests on the input rate alone. Estimated, not charged. | 0.278 measured on-GPU, no network · 0.71 est. incl. network (2x + 150 ms) | https://huggingface.co/apus-ailab/APUS-OpenJev-v1-9B |
| 74 | Decider-4B (roudone, NeoHorse 4B pointer model, author server, self-hosted) | roudone-decider-4b | 56.822253390351705 | 27.068221597395038 | 86.57628518330837 | 92.04108849933448 | 66.19900073922123 | 0.013387515151515154 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 1f4d1f7b1ce97ab27b938043444b9c2757e4a9dbecc799a803122588401f5bf1, 20 Oct additive re-freeze; this base resolves identically under the earlier 6202ef78...): base string Qwen/Qwen3.5-4B (the authors' stated Qwen3.5-4B / NeoHorse lineage, unverified) priced as Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the author server's own usage block: 662,682 input on the 1,485 answered items (equal to our offline count through the author's encode() item by item, pre-registered), 0 output tokens (one forward pass with a pointer readout, no generation); the 15 refused items count 0 tokens. 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), which this author's server refuses; their tokens count for no row's cost. | 0.048 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/roudone/Decider-4B |
| 75 | typecastlm (Mikhail Gribov, Qwen3.5-4B computed head) | typecastlm | 56.544258540063545 | 29.78900656204624 | 83.29951051808085 | 92.76354061232337 | 63.95900966831593 | 0.01589887315270936 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.030 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://github.com/mihail-gribov/typecastlm |
| 76 | RYOTIDE-Qwen (csabag, Qwen3.5-4B stock weights) | ryotide-qwen | 56.49884328343206 | 25.4219238453348 | 87.57576272152932 | 89.1284523402127 | 62.17918681512502 | 0.018226101010101012 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server (output_tokens 0, nothing generated). 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 0.018226 per 1,000 decisions (0.282x the official cost cap USD 0.064595), against USD 0.041664 over all 1,500 items (v1.6.29). | 0.096 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://github.com/csabag/ryotide |
| 77 | TypeCastLM 1.4.0 (Mikhail Gribov, Qwen3.5-4B computed head) | typecastlm-1.4.0 | 56.39414821891969 | 35.43137209354293 | 77.35692434429644 | 92.95077486551261 | 64.79986008365586 | 0.014905206499661474 | estimate | ESTIMATE (base-model reference; self-hosted open weights). NO JUDGEMENT OF OURS: pricing_v15.reference() resolves Qwen/Qwen3.5-4B straight through to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the frozen 25 Sep snapshot (the same rate aplomb-1 carries for the same base) x this row's own measured tokens. Output tokens are 0 in every answered record: one prefill, option-label readout, nothing generated. Estimated, not charged. | 0.030 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b |
| 78 | Compass 0.2.0 (adimyth, Qwen3.5-4B LoRA) | compass-0.2.0 | 56.09156024598397 | 28.917433770533798 | 83.26568672143415 | 83.28951948101809 | 57.598862966277466 | 0.025904262626262626 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server: input_tokens counts EVERY token the backbone processed including the per-candidate verification branches (2.47M over the full draw, higher than one pass by construction); output_tokens 0, nothing generated. 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 0.025904 per 1,000 decisions (0.401x the official cost cap USD 0.064595), against USD 0.049336 over all 1,500 items (v1.6.29). | 0.254 measured on-GPU, no network · 0.66 est. incl. network (2x + 150 ms) | https://huggingface.co/adimyth/compass-lora-v2 |
| 79 | Standard One 3B SH (Standard Thinking) | standard_one_3b_sh | 55.57071252242379 | 32.78619016656576 | 78.35523487828182 | 86.56241878672063 | 48.64218899335634 | 0.05151420875420875 | estimate | 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). | 0.035 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://huggingface.co/StandardThinking/StandardOne-3B-SH |
| 80 | SemIf, formerly OpenJev (Qwen3.5-4B, TheoLeeCJ) | semif-qwen3.5-4b | 55.35036926555066 | 27.10113669038628 | 83.59960184071504 | 90.76560119421038 | 63.07359395281681 | 0.01701689039408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.052 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://github.com/TheoLeeCJ/openjev |
| 81 | ArseneLupin v1.1 (IBOU-SEARCH, Qwen3.5-4B LoRA) | arsenelupin-v1.1 | 55.09132446086153 | 27.956374646954437 | 82.22627427476863 | 89.69506927060962 | 60.69086683660791 | 0.02043169696969697 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens: the server reports prompt_tokens 0 (not a measurement), so input is counted OFFLINE with the author's tokenizer and shared-prefix method (the method of the earlier measurement, measurement receipts on file; 1,011,369 over 1,485), output 0 under the 9 Oct 2026 rule (forward passes only). 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 0.020432 per 1,000 decisions (0.316x the official cost cap USD 0.064595), against USD 0.020227 over all 1,500 items (v1.6.29). | 0.084 measured on-GPU, no network · 0.32 est. incl. network (2x + 150 ms) | https://huggingface.co/IBOU-SEARCH/ArseneLupin-v1.1 |
| 82 | local-jev Qwen3.5-4B | localjev-qwen3.5-4b | 54.99411342282565 | 24.145363218466134 | 85.84286362718517 | 86.10068849774444 | 59.269627658320445 | 0.022786576354679802 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.123 measured on-GPU, no network · 0.40 est. incl. network (2x + 150 ms) | https://github.com/amithgc/local-jev |
| 83 | Decision 2B (FlyMy.AI, v59) | decision-2b | 54.82290333495299 | 22.875473256337628 | 86.77033341356835 | 90.39462801231525 | 66.37275756963365 | 0.013210160098522166 | estimate | 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 | 0.075 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://huggingface.co/flymy-ai/decision-2b-preview |
| 84 | Noma (Blackdrome AI Labs, Qwen3.5-4B-Base derivative) | noma | 54.34426226890355 | 25.051442411072966 | 83.63708212673413 | 93.34244428975083 | 61.95354260266235 | 0.01854450505050505 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-4B-Base -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens as reported by the author's server (output_tokens 0, nothing generated). 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 0.018545 per 1,000 decisions (0.287x the official cost cap USD 0.064595), against USD 0.019615 over all 1,500 items (v1.6.29). | 0.025 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/BlackdromeAILabs/noma |
| 85 | spark-s1-4b-v6 (Open Spark Jev, abhishek085) | spark-s1-4b-v6 | 54.016043277778124 | 43.2364603010071 | 64.79562625454915 | 86.95618624402076 | 60.422063196854964 | 0.020857610837438423 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 0.129 measured on-GPU, no network · 0.41 est. incl. network (2x + 150 ms) | https://github.com/abhishek085/open-spark-jev |
| 86 | Imajev-2B (mohit67890, Qwen3.5-2B LoRA) | imajev-2b | 53.96852689873508 | 21.160522003324633 | 86.77653179414553 | 81.55302418750827 | 51.76166122103225 | 0.040545683501683506 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-2B -> deepinfra:Qwen/Qwen3.5-2B USD 0.02/M input, USD 0.1/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens: input = 4-pass sum counted offline (see disclosures), output 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 0.040546 per 1,000 decisions (0.628x the official cost cap USD 0.064595), against USD 0.040140 over all 1,500 items (v1.6.29). | 0.289 measured on-GPU, no network · 0.73 est. incl. network (2x + 150 ms) | https://huggingface.co/mohit67890/imajev-2b |
| 87 | RYOTIDE-Qwen9 | ryotide_qwen9 | 53.473263564131784 | 28.676598568644657 | 78.2699285596189 | 86.71260028928383 | 49.35064060220574 | 0.048787883705206224 | estimate | Estimated exact base-model public reference: Qwen/Qwen3.5-9B Darkbloom $0.08/M input and $0.13/M output, captured 2026-10-09. Provider fp4 differs from local RYOTIDE unquantized BF16; reference tariff is not measured hosting cost. Source: https://openrouter.ai/api/v1/models/qwen/qwen3.5-9b/endpoints. | 0.132 measured on-GPU, no network · 0.41 est. incl. network (2x + 150 ms) | https://github.com/csabag/ryotide/tree/94c71a3d9f044ffaba87981e985b66f47de3be80 |
| 88 | Arbiter-4B (Goddragon, NeoHorse 4B pointer model, author server, self-hosted) | goddragon-arbiter-4b | 52.29322406130957 | 27.391250053452495 | 77.19519806916665 | 92.16491010646149 | 66.19900073922123 | 0.013387515151515154 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 1f4d1f7b1ce97ab27b938043444b9c2757e4a9dbecc799a803122588401f5bf1, 20 Oct additive re-freeze; this base resolves identically under the earlier 6202ef78...): base string Qwen/Qwen3.5-4B (the authors' stated Qwen3.5-4B / NeoHorse lineage, unverified) priced as Qwen/Qwen3.5-4B -> deepinfra:Qwen/Qwen3.5-4B USD 0.030000000000000002/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Tokens as reported by the author server's own usage block: 662,682 input on the 1,485 answered items (equal to our offline count through the author's encode() item by item, pre-registered), 0 output tokens (one forward pass with a pointer readout, no generation); the 15 refused items count 0 tokens. 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), which this author's server refuses; their tokens count for no row's cost. | 0.047 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/Goddragon/Arbiter-4B |
| 89 | ClassOne Qwen 3.5 9B (Qwen3.5-9B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted) | classone-qwen3.5-9b | 50.471173593724146 | 42.225671849648506 | 58.71667533779978 | 90.24509258883624 | 49.33529996402755 | 0.04884536222071767 | estimate | ESTIMATE (exact base-model market reference; self-hosted open weights, no per-token tariff of its own). THE ONLY ROW OF THIS RUN WHOSE RATE NEEDED NO JUDGEMENT AT ALL: Qwen/Qwen3.5-9B is in the frozen module's BASE_REFERENCES and pricing_v15.reference() resolves it straight through to qwen/qwen3.5-9b at USD 0.10 per 1M input and USD 0.15 per 1M output in the dated 25 Sep snapshot (DeepInfra's listing of the same base agrees at (0.10, 0.15)) 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. | 0.057 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/devops-thiago/classone-qwen3.5-9b |
| 90 | Liquid AI d1-3B | liquid-d1-3b | 49.50665640388703 | 22.13919485880221 | 76.87411794897184 | 94.32773191369529 | 50.56108621702079 | 0.044459420000000006 | estimate | 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 (Florian, free tiers: 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). | 0.015 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/LiquidAI/d1-3B |
| 91 | ZeroEntropy zerank-2 | zerank-2 | 49.39850940998648 | 15.596611891228248 | 83.20040692874471 | 82.89783918855927 | 48.39770052432463 | 0.05249000923645321 | estimate | 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-4B); no exact base-model floor applies | 0.175 measured on-GPU, no network · 0.50 est. incl. network (2x + 150 ms) | https://huggingface.co/zeroentropy/zerank-2-reranker |
| 92 | open-alternative-jev (Qwen3.5-4B, IkerMoel) | open-alternative-jev | 49.14497164755345 | 22.34610151162883 | 75.94384178347806 | 91.30216983336405 | 63.27859776193043 | 0.0167512315270936 | estimate | 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) | 0.043 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://github.com/ikermoel/open-alternative-jev |
| 93 | OpenSourceJev (Qwen3.5-4B Q4_K_M, native llama.cpp) | opensourcejev-qwen35-4b-q4km | 48.571422117188824 | 23.217995107845883 | 73.92484912653177 | 77.06362388237805 | 69.13657649738097 | 0.0106851539408867 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied; 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) | 0.295 measured on-GPU, no network · 0.74 est. incl. network (2x + 150 ms) | https://github.com/sabeel111/OpenSourceJev |
| 94 | 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) | exaone-4.0-1.2b-jev-v0.3 | 47.67995157039467 | 14.831288392065307 | 80.52861474872404 | 94.09940251974388 | 57.919479713067595 | 0.025274583615436694 | estimate | 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. RELEASE-LANE LINE, recorded and not acted on, AND IT IS RUN 48'S CARRIED LINE, NOT A NEW ONE: 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. | 0.020 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/carrtesy/EXAONE-4.0-1.2B-JEV |
| 95 | Malkuth-2B (newfull5, Kev post-train) | malkuth-2b | 47.633125198824516 | 15.917395195670291 | 79.34885520197874 | 92.03462875001455 | 65.5661191189412 | 0.014053873152709357 | estimate | 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 | 0.043 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://github.com/newfull5/malkuth |
| 96 | Nemotron Diffusion 8B (pst2154, optimized vLLM) | nemotron-diffusion-8b | 47.46294541124625 | 20.38848344571935 | 74.53740737677316 | 93.8305724848098 | 55.49170925809556 | 0.03045157019704434 | estimate | 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 | 0.022 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://github.com/pst2154/Nemotron_Jev |
| 97 | jiwo 0.8b (Qwen3.5-0.8B backbone with a trained decision head, one forward pass, native probabilities, self-hosted) | jiwo-0.8b | 47.20710280189251 | 15.872104931583502 | 78.54210067220151 | 94.33479301471377 | 78.40361168685378 | 0.0052466200000000004 | estimate | ESTIMATE (*est.), exact base-model market reference; self-hosted open weights with no per-token tariff of their own. The unmodified frozen pricing_v15 resolves the author's own declared base spelling Qwen/Qwen3.5-0.8B straight through to deepinfra:Qwen/Qwen3.5-0.8B at USD 0.01 per 1M input and USD 0.05 per 1M output in the dated 25 Sep 2026 snapshot. No new release-lane pricing line. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 786,993 input and 0 output tokens over 1,500 decisions. Estimated, not charged. | 0.020 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/eljiwo/jiwo-0.8b |
| 98 | decisor-4b | decisor_4b | 45.737667327238825 | 39.60641807869822 | 51.86891657577944 | 91.94556701350172 | 64.04636047691767 | 0.015792636916835696 | estimate | ESTIMATE (base-model reference; self-hosted open weights with no per-token tariff of their own). Exact base model Qwen/Qwen3.5-4B (author model card base_model, revision 1e7f195c) resolved under the v1.5-M2 base-model reference rule to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the frozen 25 Sep 2026 price cut-off snapshot (METHOD-v1.5-ADDENDUM-PRICING-INTERPRETATION-1 I-2/I-3); the listing was deprecated after the cut-off and stays the frozen reference. Output tokens priced at the reference output price. Replaces the larger-base proxy used at publication (Qwen3.5-9B Darkbloom endpoint, USD 0.08/0.13 per 1M). This row's usage on the common cost basis (1,479 items): 771,182 input and 1,479 output tokens. Estimated, not charged. | 0.048 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/underlabs/decisor-4b |
| 99 | Qwen3-Reranker-4B | qwen3-reranker-4b | 44.7457252992122 | 16.337944014530347 | 73.15350658389404 | 82.14971164581814 | 48.39770052432463 | 0.05249000923645321 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: hosted exact-model reference (deepinfra:Qwen/Qwen3-Reranker-4B); no exact base-model floor applies | 0.204 measured on-GPU, no network · 0.56 est. incl. network (2x + 150 ms) | https://huggingface.co/Qwen/Qwen3-Reranker-4B |
| 100 | decider-2b (Mapika) | decider-2b | 44.58947913615145 | 20.654576369710338 | 68.52438190259257 | 94.55945724666857 | 64.91766177740826 | 0.014771046798029557 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.015 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/Mapika/decider-2b |
| 101 | BAAI bge-reranker-v2-m3 | bge-reranker-v2-m3 | 44.08279577608597 | 0.06512222845080358 | 88.10046932372114 | 93.86976236118387 | 59.325605115070715 | 0.022688885467980296 | estimate | 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) | 0.017 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/BAAI/bge-reranker-v2-m3 |
| 102 | kev 4B (research preview) | kev-4b | 43.98178354242468 | 19.510707720813777 | 68.45285936403559 | 90.46585183268016 | 65.78036497095258 | 0.013824661330049262 | estimate | 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 | 0.072 measured on-GPU, no network · 0.29 est. incl. network (2x + 150 ms) | https://github.com/jaredpalmer/kev |
| 103 | Mixedbread mxbai-rerank-base-v2 | mxbai-rerank-base-v2 | 43.87421595931999 | 1.140707327109041 | 86.60772459153094 | 92.58758170313658 | 60.33589314238483 | 0.020996016009852216 | estimate | 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 | 0.031 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/mixedbread-ai/mxbai-rerank-base-v2 |
| 104 | lev-350m (Franck Verrot, LFM2.5-350M) | lev-350m | 43.34590358011229 | 4.827562367090417 | 81.86424479313416 | 93.98230372494703 | 80.03797760430007 | 0.00462807881773399 | estimate | 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 | 0.022 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://github.com/franckverrot/lev |
| 105 | Certo v1 (AltSlate Labs) | certo | 43.17437163074946 | 0.20278658812856143 | 86.14595667337036 | 92.53493469610669 | 96.9130801359906 | 0.0012673522167487683 | estimate | 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 | 0.042 measured on-GPU, no network · 0.23 est. incl. network (2x + 150 ms) | https://huggingface.co/altslate/certo-decision-model |
| 106 | smalljev semantic-v9 | smalljev | 42.95073007048671 | 8.670533657732655 | 77.23092648324078 | 90.21131723708834 | 60.77469204714053 | 0.020300665024630543 | estimate | 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 | 0.067 measured on-GPU, no network · 0.28 est. incl. network (2x + 150 ms) | https://github.com/isHeSatoshi/smalljev |
| 107 | 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`) | jul-fast | 42.76419280473996 | 7.4645122320968404 | 78.06387337738309 | 90.69403874154062 | 68.9870979950853 | 0.010808449559918753 | estimate | 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. RELEASE-LANE LINE, recorded and not acted on: 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. | 0.044 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/usejul/jul-decision-minicpm5-2b |
| 108 | Krino-2B | krino-2b | 42.413153726626945 | 21.99889248592056 | 62.82741496733333 | 91.85485240061512 | 61.00061799127391 | 0.019951676767676764 | estimate | v1.5-M2 base-model reference: frozen DeepInfra snapshot jevbench-v15-measure-20260925 receipts/DEEPINFRA-MODELS-20260925.json (SHA-256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f); listing Qwen/Qwen3.5-2B, USD 0.02/M input and USD 0.10/M output. The released checkpoint names Qwen/Qwen3.5-2B-Base; the frozen DeepInfra listing omits the -Base suffix. The requester cited the Qwen3.5-4B USD 0.03/M input reference, but the matching 2B family listing is used here (precedent classone-qwen3.5-2b). Self-hosted GPU deployment has no provider tariff, so price_kind is estimate. Kev generates no free-form text. However, kev/api.py output_tokens() tokenizes JSON-serialized answers, so reported usage.output_tokens is nonzero for normal answers; it is not zero 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 0.019952 per 1,000 decisions (0.309x the official cost cap USD 0.064595), against USD 0.032978 over all 1,500 items (v1.6.29). | 0.030 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/Guru0381/krino-2b |
| 109 | Alibaba GTE Reranker ModernBERT-base | gte-reranker-modernbert-base | 41.7278116825546 | 0 | 83.4556233651092 | 93.76120791185616 | 69.26965431492843 | 0.010576570197044334 | estimate | 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 | 0.024 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base |
| 110 | watt-flash-0.1 (Zaitgeist Labs, 140M mmBERT-small encoder with an option-marker scorer, one forward pass per request, self-hosted) | watt-flash-0.1 | 41.569072545871116 | 3.759262959387639 | 79.37888213235459 | 75.84832436427908 | 89.00253166733508 | 0.0023258530805687202 | mixed | 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 02 Oct deputy 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. RELEASE-LANE LINE, recorded and not acted on, AND IT IS THE CARRIED quyet/tacet LINE, NOT A NEW ONE: jhu-clsp/mmBERT-small still has no explicit market-listing decision (pricing_v15.reference() raises), so an UNLISTED_BASE_MODELS entry is the release lane's. | 0.432 measured on-CPU, no network · 1.01 est. incl. network (2x + 150 ms) | https://huggingface.co/zaitlabs/watt-flash-0.1 |
| 111 | Decision Fast (FlyMy.AI, v53a) | decision-fast | 40.59403173266273 | 6.110691078442272 | 75.07737238688318 | 91.472185793536 | 80.08651867745036 | 0.004610868226600985 | estimate | 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 | 0.058 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/flymy-ai/decision-fast-preview |
| 112 | openJev Verdict 1.4 | openjev-verdict-1.4 | 40.36740412437111 | 5.035901218168153 | 75.69890703057408 | 81.49274954261848 | 86.62313527343974 | 0.002791871921182266 | estimate | 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 | 0.293 measured on-CPU, no network · 0.74 est. incl. network (2x + 150 ms) | https://huggingface.co/heman10x/rlcd-modernbert-151m |
| 113 | Quyet-1.0-Small-EN (Chinh Nguyen, ModernBERT-base encoder with fixed heads) | quyet-1-0-small-en | 39.53807694084099 | 6.054813340422714 | 73.02134054125926 | 94.90893356488051 | 89.11296503596958 | 0.0023062222906403943 | estimate | 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 | 0.015 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/chinhnc/Quyet-1.0-Small-EN |
| 114 | ClassOne Gemma 4 E4B (Gemma-4-E4B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted) | classone-gemma4-e4b | 38.35144435275448 | 28.182424563225457 | 48.5204641422835 | 91.7664962211246 | 69.27998573050196 | 0.010568186666666667 | estimate | ESTIMATE (*est.), exact same-base market reference; self-hosted open weights with no per-token tariff of their own. pricing_v15.reference() RAISES on every spelling of google/gemma-4-E4B-it - the base is in none of the three frozen tables - so this row carries its own LABELLED estimate from the reference the dated 25 Sep 2026 snapshot itself contains for exactly this base and size, deepinfra:google/gemma-4-E4B-it at USD 0.02 per 1M input and USD 0.1 per 1M output. The already-published classone-gemma4-e2b row publishes at that same rate as a one-size-UP estimate, so for this row it is the exact size rather than an upward approximation. RELEASE-LANE LINE, recorded and not acted on: the base needs either a listing decision or an UNLISTED_BASE_MODELS entry in the frozen table; no rate was invented here and the frozen module was not modified. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 792,614 input and 0 output tokens over 1,500 decisions. Estimated, not charged. | 0.039 measured on-GPU, no network · 0.23 est. incl. network (2x + 150 ms) | https://huggingface.co/devops-thiago/classone-gemma4-e4b |
| 115 | ClassOne Qwen 3.5 4B (Qwen3.5-4B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted) | classone-qwen3.5-4b | 38.09217711089829 | 27.338167706853277 | 48.84618651494331 | 91.46618934008248 | 64.96692638610241 | 0.014715299999999999 | estimate | ESTIMATE (*est.), exact base-model market reference; self-hosted open weights with no per-token tariff of their own. The unmodified frozen pricing_v15 resolves the author's own declared base spelling Qwen/Qwen3.5-4B straight through to deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output in the dated 25 Sep 2026 snapshot. No new release-lane pricing line.; reference `deepinfra:Qwen/Qwen3.5-4B` frozen at the 25 Sep cut-off, deprecated afterwards (replaced by `Qwen/Qwen3.5-9B`); re-score if it changes (I-3) OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 735,765 input and 0 output tokens over 1,500 decisions. Estimated, not charged. | 0.043 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/devops-thiago/classone-qwen3.5-4b |
| 116 | OpenDecision (ModernBERT-large zero-shot) | opendecision | 37.99564488451158 | 7.425225288997898 | 68.56606448002526 | 87.12856375997052 | 79.10704862374335 | 0.004970862068965518 | estimate | 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 | 0.072 measured on-GPU, no network · 0.29 est. incl. network (2x + 150 ms) | https://github.com/deepanwadhwa/OpenDecision |
| 117 | Vega 4B | vegaml-4b | 37.687052549362136 | 16.689214224692094 | 58.68489087403218 | 84.39386641436894 | 47.16081979852447 | 0.057717300000000006 | estimate | v1.5-M2 base-model reference: Vega 4B is a decision engine on a frozen Qwen3.5-4B backbone (Qwen/Qwen3.5-4B, resolves in pricing_v15 to deepinfra:Qwen/Qwen3.5-4B). DeepInfra list price USD 0.03/M input, USD 0.15/M output from the dated DeepInfra catalog snapshot 2026-09-25 (jevbench-v15-measure-20260925/receipts/DEEPINFRA-MODELS-20260925.json, sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f; source https://api.deepinfra.com/models/list), read by the frozen pricing_v15 on 2026-10-10. The listing is flagged by pricing_v15 as deprecated after the 25 Sep cut-off (DEPRECATED_AFTER_CUTOFF) and stays frozen with that note. The model generates no text: 0 output tokens, cost = input tokens x USD 0.03/M. | 0.190 measured on-GPU, no network · 0.53 est. incl. network (2x + 150 ms) | https://huggingface.co/nandakishorm/vega-08b-public-intents |
| 118 | verdict-small (Manavarya09, multilingual-e5-small 118M) | verdict-small | 37.63085686322157 | 2.3417095945172277 | 72.92000413192591 | 89.48731594094681 | 100 | 0.000866152709359606 | estimate | 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 | 0.049 measured on-CPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://github.com/Manavarya09/verdict |
| 119 | kev 0.6B (research preview) | kev-0.6b | 36.93267237588606 | 7.598789169415934 | 66.26655558235618 | 91.51497362122684 | 80.09400261254245 | 0.004608220443349753 | estimate | 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 | 0.057 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://github.com/jaredpalmer/kev |
| 120 | WaterSheep (Samrat Dutta, ModernBERT-base encoder with calibrated heads, served by the author's own `watersheep --serve`) | watersheep | 35.57380209564805 | 5.305239654184986 | 65.84236453711111 | 77.5933742155104 | 68.67136467637603 | 0.011073574813811781 | estimate | 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 (02 Oct deputy rule). x this row's own measured input tokens. Estimated, not charged. No new decision. | 0.401 measured on-CPU, no network · 0.95 est. incl. network (2x + 150 ms) | https://huggingface.co/samratduttaofficial/WaterSheep |
| 121 | Raw Phi-4 mini direct logits | raw-phi-4-mini | 35.33785192837895 | 13.660237978082595 | 57.015465878675315 | 91.89255362835293 | 53.22051180195924 | 0.03625068965517241 | estimate | 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 | 0.042 measured on-GPU, no network · 0.23 est. incl. network (2x + 150 ms) | https://huggingface.co/microsoft/Phi-4-mini-instruct |
| 122 | Raw Qwen3 4B Instruct 2507 direct logits | raw-qwen3-4b-instruct-2507 | 34.56646466717713 | 33.95117742100547 | 35.18175191334879 | 91.16296741568695 | 63.36068112101963 | 0.01664602832512315 | estimate | 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 | 0.046 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507 |
| 123 | Vega 0.8B | vegaml-0.8b | 32.99469125370998 | 3.0532964296406795 | 62.93608607777928 | 89.45598874418484 | 61.47445744011434 | 0.019239100000000002 | estimate | v1.5-M2 base-model reference: Vega 0.8B is a decision engine on a frozen Qwen3.5-0.8B backbone (Qwen/Qwen3.5-0.8B, resolves in pricing_v15 to deepinfra:Qwen/Qwen3.5-0.8B). DeepInfra list price USD 0.01/M input, USD 0.05/M output from the dated DeepInfra catalog snapshot 2026-09-25 (jevbench-v15-measure-20260925/receipts/DEEPINFRA-MODELS-20260925.json, sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f; source https://api.deepinfra.com/models/list), read by the frozen pricing_v15 on 2026-10-10. The model generates no text: 0 output tokens, cost = input tokens x USD 0.01/M. | 0.085 measured on-GPU, no network · 0.32 est. incl. network (2x + 150 ms) | https://huggingface.co/nandakishorm/vega-08b-public-intents |
| 124 | kev 0.5B | kev-0.5b | 32.90230956430239 | 4.598433868646836 | 61.20618525995795 | 92.39904218127765 | 80.09398520305423 | 0.004608226600985222 | estimate | 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 | 0.044 measured on-GPU, no network · 0.24 est. incl. network (2x + 150 ms) | https://github.com/jaredpalmer/kev |
| 125 | Liquid AI d1-omni-600M | liquid-d1-omni-600m | 32.79110507364784 | 3.2615852143588597 | 62.32062493293682 | 94.0744546594334 | 73.1039158657426 | 0.00788018 | estimate | 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. | 0.021 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://huggingface.co/LiquidAI/d1-omni-600M |
| 126 | Quyet-1.0-Small (Chinh Nguyen, SEA-LION-ModernBERT-300M encoder with fixed heads) | quyet-1-0-small | 32.492363520877504 | 4.019045994532796 | 60.96568104722222 | 94.91846454857637 | 79.56261245432609 | 0.004800055418719211 | estimate | 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 | 0.015 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/chinhnc/Quyet-1.0-Small |
| 127 | Quyet-1.0-Tiny (Chinh Nguyen, mmBERT-small encoder (16 layers kept) with fixed heads) | quyet-1-0-tiny | 31.08052871088324 | 3.6964771541368573 | 58.46458026762962 | 95.09610910731021 | 88.43590631389968 | 0.0024292364532019703 | estimate | 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 | 0.013 measured on-GPU, no network · 0.18 est. incl. network (2x + 150 ms) | https://huggingface.co/chinhnc/Quyet-1.0-Tiny |
| 128 | Open Jev JSON Canvas (JoshuaSP) | open-jev-json-canvas-joshuasp | 30.596109035139946 | 61.19221807027989 | 0 | 85.53332924416175 | 49.295219536461964 | 0.04899585591133005 | estimate | 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 | 0.148 measured on-GPU, no network · 0.45 est. incl. network (2x + 150 ms) | https://github.com/JoshuaSP/open-jev |
| 129 | Phocinae-Largha-150M-v1 | phocinae-largha-150m-v1 | 30.240174277642033 | 3.609767391262904 | 56.87058116402116 | 77.11764018052709 | 95.30667675976518 | 0.0014336531986531987 | estimate | 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 no new release-lane pricing line is created. 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). | 0.443 measured on-CPU, no network · 1.04 est. incl. network (2x + 150 ms) | https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1 |
| 130 | openJev Verdict (heman10x, ModernBERT-base 151M) | openjev-verdict | 26.63220276955003 | 4.1496399816416485 | 49.11476555745841 | 79.40046783928 | 86.62313527343974 | 0.002791871921182266 | estimate | 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 | 0.269 measured on-CPU, no network · 0.69 est. incl. network (2x + 150 ms) | https://github.com/Heman10x-NGU/openJev-verdict-2.0 |
| 131 | Tacet Sonata (CodePawl, 144M packed encoder on mmBERT-small, option-marker softmax in one pass, in-process via the author's tacet package) | tacet-sonata | 24.310999978028352 | 2.2436273530567146 | 46.37837260299999 | 80.28211114890141 | 89.01467380421724 | 0.0023236865267433986 | estimate | 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 02 Oct deputy rule for classifier read-outs. RELEASE-LANE LINE, recorded and not acted on, AND IT IS THE QUYET ROWS' CARRIED LINE, NOT A NEW ONE: jhu-clsp/mmBERT-small still has no explicit market-listing decision (pricing_v15.reference() raises), so an UNLISTED_BASE_MODELS entry is the release lane's. | 0.231 measured on-CPU, no network · 0.61 est. incl. network (2x + 150 ms) | https://huggingface.co/codepawl/tacet-sonata |
| 132 | CLM-8B (Contrastive-LM, clm-latest) | clm-8b | 20.086328406467768 | 0.13075619159747265 | 40.041900621338065 | 92.63548383949004 | 50.501662240594015 | 0.044662660714285714 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.021 measured on-GPU, no network · 0.19 est. incl. network (2x + 150 ms) | https://github.com/Contrastive-LM/CLM |
| 133 | ClassOne Qwen 3.5 2B (Qwen3.5-2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted) | classone-qwen3.5-2b | 19.21888869011457 | 5.814527660927306 | 32.62324971930184 | 93.11279379161611 | 70.24966415777286 | 0.009810200000000002 | estimate | ESTIMATE (*est.), exact base-model market reference; self-hosted open weights with no per-token tariff of their own. The unmodified frozen pricing_v15 resolves the author's own declared base spelling Qwen/Qwen3.5-2B straight through to deepinfra:Qwen/Qwen3.5-2B at USD 0.02 per 1M input and USD 0.1 per 1M output in the dated 25 Sep 2026 snapshot. No new release-lane pricing line. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 735,765 input and 0 output tokens over 1,500 decisions. Estimated, not charged. | 0.030 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/devops-thiago/classone-qwen3.5-2b |
| 134 | Deem 0.8B v1 | deem-0.8-v1 | 18.07951706793502 | 6.531204174677637 | 29.6278299611924 | 85.44505463939521 | 80.28784730320886 | 0.004540166256157636 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: documented hosted-model estimate. Frozen market reference deepinfra:Qwen/Qwen3.5-0.8B. | 0.157 measured on-GPU, no network · 0.46 est. incl. network (2x + 150 ms) | https://huggingface.co/LibertAIDAI/deem-0.8-v1 |
| 135 | Laya multilingual | laya-multilingual | 16.971882566936873 | 0.2879184668737643 | 33.655846666999985 | 77.82044954853706 | 82.1551651086044 | 0.003933940886699508 | estimate | 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. | 0.358 measured on-CPU, no network · 0.87 est. incl. network (2x + 150 ms) | https://huggingface.co/convaiinnovations/laya/tree/1c5edc17a7acd8701df6fc341c0d179f1c62c982/multilingual |
| 136 | ClassOne Gemma 4 E2B (Gemma-4-E2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted) | classone-gemma4-e2b | 15.208879168420996 | 7.802105373202749 | 22.615652963639242 | 93.2878888255429 | 70.20313863040258 | 0.00984529451591063 | estimate | ESTIMATE (same-family reference, one size up; self-hosted open weights, no bookable listing for this base). pricing_v15.reference() RAISES on google/gemma-4-E2B and on google/gemma-4-e2b - the base is in none of the three frozen tables - and the published page carries no priced row on this base either: the one Gemma-4-E2B entry on the board, system-one-open, is listed not_measured. The row therefore carries its own labelled estimate from the nearest reference the frozen 25 Sep snapshot itself contains, deepinfra:google/gemma-4-E4B-it at USD 0.02 per 1M input and USD 0.10 per 1M output - the SAME Gemma-4 effective-parameter family, one size LARGER, so it is both the nearest available reference and the direction that cannot flatter this row - 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. RELEASE-LANE LINE, recorded and not acted on: google/gemma-4-E2B needs either a listing decision or an UNLISTED_BASE_MODELS entry in the frozen table before this row may be published; receipts/COST-ALTERNATIVES-R58.json has the alternatives' arithmetic already done. | 0.029 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/devops-thiago/classone-gemma4-e2b |
| 137 | Julia-1 (Supersonic Labs) | julia-1 | 9.854654871580568 | 2.4412732607743255 | 17.26803648238681 | 79.42119248319958 | 89.74859303626371 | 0.0021964107744107744 | estimate | 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: CoS deputy decision #18761 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). Florian's standing 24 Sep 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 no new release-lane pricing line is created. 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). | 0.292 measured on-CPU, no network · 0.73 est. incl. network (2x + 150 ms) | https://huggingface.co/SupersonicLabs/Julia-1 |
| 138 | Raw Qwen3 1.7B direct logits | raw-qwen3-1.7b | 8.320306683538828 | 5.124286554530066 | 11.516326812547591 | 93.23530725162705 | 68.54983190559568 | 0.011177352216748768 | estimate | 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 | 0.032 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/Qwen/Qwen3-1.7B |
| 139 | Raw Qwen3 0.6B direct logits | raw-qwen3-0.6b | 8.247746904337834 | 5.554864880982027 | 10.94062892769364 | 93.52703208725202 | 77.58073177551512 | 0.005588676108374384 | estimate | 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 | 0.029 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/Qwen/Qwen3-0.6B |
| outside caps | Respan Span-02 | respan-span-02 | 69.33905813017047 | 60.897967232257315 | 77.78014902808363 | 85.32322293910049 | 45.07530479505206 | 0.06773656565656565 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff for these weights): the frozen market reference google/gemma-4-26b-a4b-it USD 0.09/M input, USD 0.3/M output for the declared base Gemma 4 26B-A4B (catalog id google/gemma-4-26b-a4b-it, supplied by the evaluator because the frozen path matches base names case-sensitively; the submission's model card spells it google/gemma-4-26B-A4B-it) x this row's own measured tokens on the common 1,485-item basis (776,560 input and 102,328 output tokens as reported by the author's server; the 15 refused items count as zero). Estimated, not charged; not a quote from the author or any host. OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.067737 per 1,000 decisions on the common basis (USD 0.067059 over all 1,500 items) is 1.049x 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 0.067737 per 1,000 decisions (1.049x the official cost cap USD 0.064595), against USD 0.067059 over all 1,500 items (v1.6.29). | 0.027 measured on-GPU, no network · 0.20 est. incl. network (2x + 150 ms) | https://huggingface.co/lihe07/span-02 |
| outside caps | Chon-1.0-E4B | chon-1.0-e4b | 68.39271102371643 | 54.96148352069754 | 81.82393852673533 | 43.95627533510597 | 70.74636118712607 | 0.009443245791245792 | listed | Exact listed rate of the base model google/gemma-4-E4B-it (DeepInfra, frozen 25 Sep 2026 catalog snapshot, USD 0.02/M input, 0.10/M output); this line is added to the frozen pricing module at its next re-freeze. No output tokens under the author's --no-think mode. COMMON COST BASIS: the published cost uses the same remaining 1,485 items as every other row: 701,161 input tokens, USD 0.009443245791245792 per 1,000 decisions (0.146x cap). ALL-ITEMS DIAGNOSTIC: USD 0.010571213333333333 per 1,000 over all 1,500 items (0.164x cap). All items still count for Intelligence. | 22.177 measured on-CPU, no network · 44.50 est. incl. network (2x + 150 ms) | https://huggingface.co/ninhnv/chon-1.0-e4b |
| outside caps | swanOne (blockbrain, Qwen3.8-Flash-Next NVFP4) | swanone | 72.36627654122822 | 56.15104294255194 | 88.58151013990448 | 82.64878475169937 | 42.15074720483203 | 0.08478293103448274 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.281 measured on-GPU, no network · 0.71 est. incl. network (2x + 150 ms) | https://github.com/blockbrain-ai/swanone-recipe |
| outside caps | NInfer Qwen3.8-Flash-Next mixed | ninfer-qwen3.8-flash-next | 69.67631278082561 | 48.980301026876646 | 90.37232453477458 | 89.33173525631216 | 42.53362878447125 | 0.08232764778325123 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.080 measured on-GPU, no network · 0.31 est. incl. network (2x + 150 ms) | https://github.com/igorls/ninfer |
| outside caps | Qwen3.5-9B Jev-like data-mix v2 | qwen35-9b-jev-data-mix-v2 | 60.86540021228136 | 41.94184606895763 | 79.78895435560509 | 85.40793837616167 | 45.690218065856605 | 0.06461391625615764 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.168 measured on-GPU, no network · 0.49 est. incl. network (2x + 150 ms) | https://huggingface.co/jsaurabh/qwen3.5-9b-jev-data-mix-v2 |
| outside caps | JevOne | jevone | 68.8604478570867 | 45.955077428728934 | 91.76581828544445 | 89.83885080776159 | 39.84212731591276 | 0.10121908866995075 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.062 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/juspay/jev-one |
| outside caps | JPT-35B-A3B (Qwen3.5-35B-A3B fine-tune) | jpt-35b-a3b | 70.67673257774709 | 51.42208418374838 | 89.9313809717458 | 90.51950606084544 | 33.627578514713974 | 0.16308501184834123 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): OpenRouter qwen/qwen3.5-35b-a3b in the frozen 25 Sep snapshot (exact declared base; no explicit listing-decision line in the frozen module: release-lane line), USD 0.3125 per 1M input and USD 1.25 per 1M output x this row's own measured tokens. Estimated, not charged | 0.071 measured on-GPU, no network · 0.29 est. incl. network (2x + 150 ms) | https://huggingface.co/kirp/jpt-35b-a3b |
| outside caps | decider-35b-a3b (Mapika) | decider-35b-a3b | 66.1715118587882 | 48.72471958402083 | 83.61830413355554 | 91.27357492980725 | 34.38579876859529 | 0.1538650708128079 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.050 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://huggingface.co/Mapika/decider-35b-a3b |
| outside caps | Verdin-4B (Victor Bennett, Qwen3.5-4B + merged LoRA r16 + pointer head) | verdin-4b | 59.540149828182294 | 35.50482714952485 | 83.57547250683973 | 45.8869739530126 | 62.17924458049355 | 0.018226020202020205 | estimate | ESTIMATE (*est.) through the frozen pricing_v15 path: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0, the same reference). The base model Qwen/Qwen3.5-4B carries an EXACT base-model market reference in the frozen 25 Sep catalog - deepinfra:Qwen/Qwen3.5-4B (snapshot sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f) - so resolve() returns rankable:true with a labelled ESTIMATE and no new pricing line is needed. The operator tariff is excluded by the 30-day rule, and the reference itself is flagged in the module as deprecated after the cut-off (replaced by Qwen/Qwen3.5-9B): re-score if it changes (I-3). Output tokens are a literal 0 for every item - the model generates no text at all, and the author's own server reports usage.output_tokens: 0 (the pjev-2b convention and the 9 Oct 2026 rule for native option-probability models) - so the output term is 0.0, not 0.15. Multiplied by this row's own measured input tokens (902,188 input tokens; cost per 1,000 decisions over the common 1,485-item basis). 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 0.018226 per 1,000 decisions (0.282x the official cost cap USD 0.064595), against USD 0.018044 over all 1,500 items (v1.6.29). | 17.662 measured on-CPU, no network · 35.47 est. incl. network (2x + 150 ms) | https://huggingface.co/victorbennett/Verdin-4B |
| outside caps | StartLux-Decision v1.1 27B Instant (StartLux Labs, Qwen3.8-27B fine-tune) | startlux-decision-v1.1-27b-instant | 82.55772091558464 | 73.19432958384127 | 91.92111224732803 | 90.62312905261143 | 30.41280058663709 | 0.20872444444444446 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 3948e3f1...ea4d0, the module this row was first priced with (the later re-freeze only adds a hosted-API launch-list key), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). The model cards do not name the Qwen version; Qwen/Qwen3.8-27B is the base the Decision Index metadata gives for this model, consistent with the NOTICE (modified Alibaba Cloud Apache-2.0 model) and with the config (Qwen3_5ForConditionalGeneration, 64 layers, hidden 5120), the same reference as the live torchcast-decision-27b row (a fine-tune of StartLux-Decision-27B v1). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.208724 per 1,000 decisions on the common basis (USD 0.206637 over all 1,500 items) is 3.231x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b 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 0.208724 per 1,000 decisions (3.231x the official cost cap USD 0.064595), against USD 0.206637 over all 1,500 items (v1.6.29). | 0.055 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-Instant |
| outside caps | RSI-Jev v6.1-VL 27B | rsi-jev-v6.1-vl-27b | 73.45697484423279 | 61.564156247760415 | 85.34979344070516 | 84.71397705257328 | 31.141652902933714 | 0.19736867999999996 | estimate | ESTIMATE (*est.), base-model market reference. The declared base Qwen/Qwen3.8-27B resolves in the frozen 25 Sep 2026 snapshot through pricing_v15 BASE_REFERENCES to qwen/qwen3.8-27b at USD 0.42 per 1M input and USD 3.00 per 1M output (verified before the pod was rented; r75 COST-BASIS.json). This row reports a real output token count - 1 token per readout, 1,471 in total - so the Gutsy zero-output rule does NOT apply and the output price is charged as measured. This row's own usage on draw v1.6-regular-20261010-a2: 694,381 input and 1,471 output tokens over 1,500 decisions. | 0.123 measured on-GPU, no network · 0.40 est. incl. network (2x + 150 ms) | https://huggingface.co/shgao/rsi-jev-v6.1-vl-27b |
| outside caps | JEV-27B (AutoTrust, Qwen3.8-27B) | jev-27b-autotrust | 73.86417008485257 | 56.828664559360675 | 90.89967561034447 | 86.89503266505372 | 31.02571003473804 | 0.19913289099526063 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.105 measured on-GPU, no network · 0.36 est. incl. network (2x + 150 ms) | https://huggingface.co/autotrust/JEV-27B |
| outside caps | torchcast-decision-27b (Torchcast AI, Qwen3.8-27B fine-tune) | torchcast-decision-27b | 81.56565836114741 | 71.75187578095783 | 91.379440941337 | 87.72436167205926 | 30.341044859222123 | 0.20987715639810428 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.097 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://huggingface.co/torchcast-ai/torchcast-decision-27b |
| outside caps | Bespoke Nimble 9B (Bespoke Labs) | nimble-9b | 56.514998578191054 | 43.14237010986392 | 69.88762704651819 | 86.11327010751721 | 36.75037423027614 | 0.1283278325123153 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.145 measured on-GPU, no network · 0.44 est. incl. network (2x + 150 ms) | https://github.com/bespokelabsai/nimble |
| outside caps | pplx-decider-v1.2 (Perplexity, Qwen3.8-27B full fine-tune, marker readout, self-hosted) | pplx-decider-v1.2 | 81.74185907802455 | 72.71358625871065 | 90.77013189733844 | 86.32357007699105 | 30.0417413123765 | 0.2147543434343434 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 6202ef78a2c83cbffc9389cf1f37df70fe0c85ed03f4dd5c5f997f3d31aae13a): declared base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output. Estimated, not charged, not a quote from the author or any host. USD per 1,000 decisions on the denominator of the release basis (the same-items cost rule for this draw prices the common item set without the 15 very long items, for every row alike). Same pair as the published pplx-decider-v1-1-27b and vllm-sr-d3 rows. Tokens as reported by the author's own `Decider.decide()` usage block (input = every non-padding position the request used; output 0, nothing is generated - the logits are read at the markers in one forward pass). The offline count through the author's own protocol.Encoder predicted the server's numbers EXACTLY, item by item: 1,939,205 input tokens over all 1,500 answered items, and the same prediction held on the 231 public items (154,452) before any sealed item was dispatched. OVER THE OFFICIAL COST CAP on the common 1,485-item basis: USD 0.214754343 per 1,000 decisions (3.325x cap); the row stays RANKED, but is outside the official cost-capped views and visible in uncapped views. ALL-ITEMS DIAGNOSTIC: USD 0.542977 per 1,000 over all 1,500 answered items (8.41x cap). All 1,500 items count for Intelligence. | 0.142 measured on-GPU, no network · 0.43 est. incl. network (2x + 150 ms) | https://huggingface.co/perplexity-ai/pplx-decider-v1.2 |
| outside caps | reflex 4B (kshetrajna12) | reflex-4b | 58.95740703790636 | 30.935558296483705 | 86.97925577932901 | 68.7900017304639 | 63.142712976263866 | 0.016926853448275863 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate; 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) | 1.358 measured on-GPU, no network · 2.87 est. incl. network (2x + 150 ms) | https://github.com/kshetrajna12/reflex |
| outside caps | Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout) | pplx-decider-v1-1-27b | 82.79186839726479 | 75.71677945512734 | 89.86695733940225 | 86.79279341461623 | 29.886617009940878 | 0.21732654028436021 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.122 measured on-GPU, no network · 0.39 est. incl. network (2x + 150 ms) | https://huggingface.co/perplexity-ai/pplx-decider-v1.1-27b |
| outside caps | vLLM-SR d3 | vllm-sr-d3 | 82.96592796157353 | 73.73157285955872 | 92.20028306358832 | 87.36893361347018 | 29.839473214267542 | 0.21811434343434347 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff for these weights): the frozen Qwen/Qwen3.8-27B market reference of the published v1.6 field, USD 0.42 per 1M input and USD 3.00 per 1M output - the same pair the published pplx-decider-v1-1-27b and torchcast-decision-27b rows carry - x this row's own measured tokens on the common 1,485-item basis (1,951,205 input tokens; the author's server reports 0 output tokens because nothing is generated). USD 0.218114 per 1,000 decisions on the common basis (USD 0.5463374 over all 1,500 items), about 3.4x the published v1.6 ceiling of USD 0.0646 per 1,000 decisions, so this row is outside the default cost-capped Capability ranking and the squared cost gate keeps its composite small. Estimated, not charged; not a quote from the author or any host. 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 0.218114 per 1,000 decisions (3.377x the official cost cap USD 0.064595), against USD 0.546337 over all 1,500 items (v1.6.29). | 0.109 measured on-GPU, no network · 0.37 est. incl. network (2x + 150 ms) | https://huggingface.co/vllm-sr/d3 |
| outside caps | vLLM-SR d3-turbo | vllm-sr-d3-turbo | 82.59113932461057 | 74.82897422466625 | 90.35330442455489 | 86.03728492328564 | 29.839473214267542 | 0.21811434343434347 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff for these weights): declared base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output - the same pair the published vllm-sr-d3 row on this draw carries - x this row's own measured tokens on the common 1,485-item basis (771,190 input tokens on the 1,485 answered items; the 15 items over the author's 32,768-token default cap are refused by the author's own server with 0 tokens counted; output 0 because nothing is generated). Estimated, not charged; not a quote from the author or any host. OVER THE OFFICIAL COST CAP: USD 0.21811434343434347 per 1,000 decisions on the 1,485 common items, 3.38x the official cap USD 0.06459465517241379; ranked and disclosed following the published vllm-sr-d3 precedent. The frozen Qwen3.8-27B pair is also carried by published startlux-decision-v1.1-27b-instant and startlux-decision-v1.1-27b-autothink. ALL-ITEMS DIAGNOSTIC: USD 0.2159332 per 1,000 decisions over the full 1,500-item denominator, 3.34x the official cap; the published cost above uses the same remaining 1,485 items as every other row. All items still count for Intelligence. | 0.144 measured on-GPU, no network · 0.44 est. incl. network (2x + 150 ms) | https://huggingface.co/vllm-sr/d3-turbo |
| outside caps | JADE (Qwen3.8-27B LoRA) | jade | 49.24946957396523 | 47.93439876939806 | 50.5645403785324 | 85.53222225487117 | 33.57969498589053 | 0.16368548408937036 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.127 measured on-GPU, no network · 0.40 est. incl. network (2x + 150 ms) | https://huggingface.co/theunnecessarythings/JADE |
| outside caps | Nagi-ENORMOUS 27B (nagisanzenin, Qwen3.8-27B LoRA) | nagi-enormous | 73.76536425253485 | 62.38818589248167 | 85.14254261258804 | 86.16105911802293 | 29.98821307073638 | 0.21563846464646463 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the evaluator's wrapper (input = attention-mask ones over every forward pass, output_tokens 0). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.215638 per 1,000 decisions on the common basis (USD 0.213482 over all 1,500 items) is 3.338x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b 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 0.215638 per 1,000 decisions (3.338x the official cost cap USD 0.064595), against USD 0.213482 over all 1,500 items (v1.6.29). | 0.151 measured on-GPU, no network · 0.45 est. incl. network (2x + 150 ms) | https://huggingface.co/nagisanzeninz/Nagi-ENORMOUS |
| outside caps | Jebadiah 27B (Frontier Infra, Qwen3.8-27B LoRA) | jebadiah-27b | 70.3119788331729 | 57.58551644935426 | 83.03844121699153 | 85.83700487110255 | 29.935481526523958 | 0.2165129857819905 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.139 measured on-GPU, no network · 0.43 est. incl. network (2x + 150 ms) | https://huggingface.co/frontier-infra/jebadiah-27b |
| outside caps | Decision 2.0 Vega 27B (vLLM-SR, Qwen3.8-27B) | decision-2-0-vega-27b | 71.30534158412956 | 60.4093162408047 | 82.20136692745442 | 83.80610555916236 | 29.632311028770005 | 0.22161014218009478 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.190 measured on-GPU, no network · 0.53 est. incl. network (2x + 150 ms) | https://huggingface.co/vllm-sr/Decision-2.0-Vega-27B |
| outside caps | Standard One 8B SH (Standard Thinking) | standard_one_8b_sh | 59.981611956556755 | 34.464905061466695 | 85.49831885164681 | 84.28099014807135 | 43.21124696592469 | 0.0781553 | estimate | 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. | 0.061 measured on-GPU, no network · 0.27 est. incl. network (2x + 150 ms) | https://huggingface.co/StandardThinking/StandardOne-8B-SH |
| outside caps | AutoJev-27B (denis-pplx, Qwen3.8-27B) | autojev-27b | 71.61983491755967 | 55.62422706919757 | 87.61544276592178 | 88.60757446254297 | 29.365937020749882 | 0.22618758620689655 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.097 measured on-GPU, no network · 0.34 est. incl. network (2x + 150 ms) | https://huggingface.co/denis-pplx/autojev-27b |
| outside caps | Eikos-27B (caiovicentino1, Qwen3.8-27B) | eikos-27b | 77.18992599407068 | 64.40270761245412 | 89.97714437568723 | 88.08630815059152 | 28.689752882469676 | 0.23823646551724137 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.113 measured on-GPU, no network · 0.38 est. incl. network (2x + 150 ms) | https://huggingface.co/caiovicentino1/eikos-27b |
| outside caps | NInfer Qwen3.8-27B NVFP4 | ninfer-qwen3.8-27b | 67.99284998901135 | 51.800126951055375 | 84.18557302696732 | 90.47734239273063 | 29.118887844204664 | 0.2305174137931034 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.055 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://github.com/igorls/ninfer |
| outside caps | JEB-35B-A3B (szybkie.ai, Qwen3.6-35B-A3B LoRA) | jeb-35b-a3b | 61.6062187771955 | 33.22856158102424 | 89.98387597336676 | 92.99325391268877 | 42.92470023353102 | 0.07989323232323232 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.6-35B-A3B -> qwen/qwen3.6-35b-a3b USD 0.15/M input, USD 1.0/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused or failed items cost nothing and still count as wrong answers). Tokens as reported by the author's server, KEPT AS SENT: output_tokens is 1 per answered question (the label position the engine counts, not generated text); this adds about USD 0.0002 per 1,000 decisions at the reference output rate. OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.079893 per 1,000 decisions on the common basis (USD 0.079094 over all 1,500 items) is 1.237x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b 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 0.079893 per 1,000 decisions (1.237x the official cost cap USD 0.064595), against USD 0.079094 over all 1,500 items (v1.6.29). | 0.035 measured on-GPU, no network · 0.22 est. incl. network (2x + 150 ms) | https://huggingface.co/szybkie-ai/jeb-35b-a3b |
| outside caps | OpenJev (thinking, BF16) | openjev-thinking | 75.87983429229429 | 70.34621384425424 | 81.41345474033434 | 72.55270827201397 | 28.51480924396573 | 0.24145693965517237 | estimate | 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 | 0.857 measured on-GPU, no network · 1.86 est. incl. network (2x + 150 ms) | https://github.com/razorback16/openjev |
| outside caps | Open-Jev 9B (Zefan Cai) | open-jev-zefan-9b | 61.84549459243814 | 43.88591848753404 | 79.80507069734225 | 74.39486349899005 | 33.05490478272432 | 0.17041317733990147 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.618 measured on-GPU, no network · 1.39 est. incl. network (2x + 150 ms) | https://github.com/Zefan-Cai/Open-Jev |
| outside caps | Standard One 8B (Standard Thinking) | standardone-8b | 58.40967348707115 | 32.80426604507879 | 84.01508092906352 | 92.76442849252426 | 43.283250331904235 | 0.07772456896551724 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.032 measured on-GPU, no network · 0.21 est. incl. network (2x + 150 ms) | https://huggingface.co/StandardThinking/StandardOne-8B |
| outside caps | Clef (Cloudflare, Qwen3.8-27B post-train with a joint schema head, multimodal, measured on text) | clef | 75.32934491938597 | 59.08165501447566 | 91.57703482429629 | 83.3998209667487 | 28.102157220114762 | 0.24922681034482758 | estimate | latest published v1.5 row (v1.5.8 combined file) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied; self-hosted estimate, not measured hosted billing | 0.213 measured on-GPU, no network · 0.58 est. incl. network (2x + 150 ms) | https://huggingface.co/Cloudflare/clef |
| outside caps | pjev-2b (Kailune AI, frozen MiniCPM5-2B + merged LoRA r64, letter-logit readout, T=1 stub) | pjev-2b | 57.71754900661349 | 28.96851828460498 | 86.466579728622 | 49.10963929819891 | 63.52461077797332 | 0.01643789898989899 | estimate | 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. No new release-lane pricing line is created. 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). | 10.063 measured on-CPU, no network · 20.28 est. incl. network (2x + 150 ms) | https://huggingface.co/Kailune-AI/pjev-2b |
| outside caps | Bev / Bonsai 27B | bev-bonsai-27b | 66.13873250860377 | 43.77513601921869 | 88.50232899798883 | 72.98222461463935 | 28.231834531826323 | 0.2467585344827586 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE: documented hosted-model estimate. Frozen market reference qwen/qwen3.8-27b. | 0.986 measured on-GPU, no network · 2.12 est. incl. network (2x + 150 ms) | https://huggingface.co/Reza2kn/Bev |
| outside caps | LitJev (Qwen3.8-27B) | litjev | 66.60329337630368 | 43.28549422942267 | 89.9210925231847 | 68.17671225113584 | 28.35809912639418 | 0.2443787068965517 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 1.473 measured on-GPU, no network · 3.10 est. incl. network (2x + 150 ms) | https://github.com/zhengxuyu/litjev |
| outside caps | system-one (Qwen3-8B, Sean Goedecke) | system-one-sg | 29.718120999285617 | 29.271508779158296 | 30.16473321941294 | 90.8867694864598 | 44.96855803039237 | 0.06829381896551724 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.050 measured on-GPU, no network · 0.25 est. incl. network (2x + 150 ms) | https://github.com/sgoedecke/system-one |
| outside caps | StartLux-Decision v1.1 27B AutoThink (StartLux Labs, Qwen3.8-27B fine-tune, gate 0.6) | startlux-decision-v1.1-27b-autothink | 80.34729259324169 | 68.96369650479022 | 91.73088868169317 | 75.05228443363347 | 23.51863513364647 | 0.35430626262626264 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 3948e3f1...ea4d0, the module this row was first priced with (the later re-freeze only adds a hosted-API launch-list key), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). The model cards do not name the Qwen version; Qwen/Qwen3.8-27B is the base the Decision Index metadata gives for the sibling Instant variant (AutoThink itself is not on the Decision Index), shares the same NOTICE, consistent with the NOTICE (modified Alibaba Cloud Apache-2.0 model) and with the config (Qwen3_5ForConditionalGeneration, 64 layers, hidden 5120), the same reference as the live torchcast-decision-27b row (a fine-tune of StartLux-Decision-27B v1). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.354306 per 1,000 decisions on the common basis (USD 0.350763 over all 1,500 items) is 5.485x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b 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 0.354306 per 1,000 decisions (5.485x the official cost cap USD 0.064595), against USD 0.350763 over all 1,500 items (v1.6.29). | 0.056 measured on-GPU, no network · 0.26 est. incl. network (2x + 150 ms) | https://huggingface.co/startlux-models/StartLux-Decision-v1.1-27B-AutoThink |
| outside caps | kev 8B (research preview) | kev-8b | 50.0091761607623 | 29.26217941280341 | 70.7561729087212 | 88.97143473610706 | 40.42054386073457 | 0.09682374384236453 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): price floor: base-model reference price applied | 0.074 measured on-GPU, no network · 0.30 est. incl. network (2x + 150 ms) | https://github.com/jaredpalmer/kev |
| outside caps | Imajev-9B (mohit67890, Qwen3.5-9B LoRA) | imajev-9b | 62.42104465724772 | 36.7327337780622 | 88.10935553643324 | 79.48842001125843 | 30.792561090951693 | 0.20272841750841747 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base Qwen/Qwen3.5-9B -> qwen/qwen3.5-9b USD 0.09999999999999999/M input, USD 0.15/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835, the module this row was first priced with (the later re-freezes only add hosted-API launch-list keys), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). Tokens: input = 4-pass sum counted offline (see disclosures), output 0. OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.202728 per 1,000 decisions on the common basis (USD 0.200701 over all 1,500 items) is 3.138x 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 0.202728 per 1,000 decisions (3.138x the official cost cap USD 0.064595), against USD 0.200701 over all 1,500 items (v1.6.29). | 0.383 measured on-GPU, no network · 0.92 est. incl. network (2x + 150 ms) | https://huggingface.co/mohit67890/imajev-9b |
| outside caps | Raw Qwen3 8B direct logits | raw-qwen3-8b | 30.13533882926388 | 26.54912920590792 | 33.721548452619835 | 88.65282990878164 | 45.53515592313027 | 0.0653875104679803 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.078 measured on-GPU, no network · 0.31 est. incl. network (2x + 150 ms) | https://huggingface.co/Qwen/Qwen3-8B |
| outside caps | Nev-2B (EryriLabs) | nev-2b | 49.91717019137964 | 20.23194846537782 | 79.60239191738147 | 54.82852093556197 | 71.03778132505249 | 0.00923437037037037 | estimate | ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module (sha256 dcfcecb0..., the module in force; its base-model references are unchanged since the dae86740... module this row was first priced with - the two later re-freezes each added one hosted-API launch-list key and nothing else): declared base Qwen/Qwen3.5-2B-Base -> deepinfra:Qwen/Qwen3.5-2B USD 0.02/M input, USD 0.1/M output. The author's shim reports usage {calls, permutations, backend} and NO token fields at all, so the input tokens are counted OFFLINE with the author's own prompt code (nev.layout.single_prompts over hev.decider_layout.render_row) at the exact serving settings, and cross-checked against llama-server's own prompt-token counter delta from the measured pass; output tokens are 0, because one readout position per call is restricted to the option letters and nothing is returned as text (the pjev-2b convention and the 9 Oct CoS rule for native option-probability models). Estimated, not charged, and not a quote from the author. 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 0.009234 per 1,000 decisions (0.143x the official cost cap USD 0.064595), against USD 0.009142 over all 1,500 items (v1.6.29). | 6.343 measured on-CPU, no network · 12.84 est. incl. network (2x + 150 ms) | https://huggingface.co/EryriLabs/Nev-2B-GGUF |
| outside caps | Clef-omni | clef-omni | 58.315926732426384 | 34.47716016566848 | 82.15469329918429 | 86.7880661703893 | 26.792900075221908 | 0.275573 | estimate | ESTIMATE (*est.), base-class reference. Self-hosted open weights; the declared base Qwen/Qwen3-Omni-30B-A3B-Instruct is not a key of any frozen pricing table. Method as the v1.6.5 Liquid rows: the nearest same-family entry of the frozen 25 Sep 2026 snapshot, deepinfra:Qwen/Qwen3-VL-30B-A3B-Instruct (the multimodal Qwen3 30B-A3B), at USD 0.15 per 1M input and USD 0.60 per 1M output. The authors' systemone() has no generation path, so output_tokens is a literal 0. This row's own usage on draw v1.6-regular-20261010-a2: 2,755,730 input and 0 output tokens over 1,500 decisions. | 0.143 measured on-GPU, no network · 0.44 est. incl. network (2x + 150 ms) | https://huggingface.co/Cloudflare/clef-omni |
| outside caps | Kev 27B (Kev 1.0, Qwen3.8-27B) | kev-27b | 78.93896122956639 | 64.16071886717361 | 93.71720359195915 | 83.65684843680415 | 18.302044660887503 | 0.5287680162491537 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.180 measured on-GPU, no network · 0.51 est. incl. network (2x + 150 ms) | https://huggingface.co/jaredpalmer/kev-27b |
| outside caps | SimpleJev Qwen3.8-27B (self-hosted, v1.6 pool) | simplejev-qwen3.8-27b-selfhost | 74.89270871895144 | 57.92931033189586 | 91.85610710600702 | 76.19964460590472 | 15.10193161538902 | 0.6759827488151658 | estimate | ESTIMATE (base-model reference; self-hosted open weights, no public tariff): frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price, the pair the published autojev-27b, eikos-27b and open-jev-zefan-27b-v1.1 rows use, USD 0.42 per 1M input and USD 3.0 per 1M output x this row's own measured tokens. Estimated, not charged | 0.622 measured on-GPU, no network · 1.39 est. incl. network (2x + 150 ms) | https://github.com/featherless-ai/simple-jev |
| outside caps | Open-Jev 2B (Zefan Cai) | open-jev-zefan-2b | 47.71137138509946 | 21.804441188912403 | 73.61830158128652 | 77.13731150142542 | 33.05490478272432 | 0.17041317733990147 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 0.437 measured on-GPU, no network · 1.02 est. incl. network (2x + 150 ms) | https://github.com/Zefan-Cai/Open-Jev |
| outside caps | Open-Jev 27B v1.1 (Zefan Cai, LoRA + decision head on Qwen3.8-27B) | open-jev-zefan-27b-v1.1 | 67.96306879751505 | 52.45777412397041 | 83.46836347105969 | 72.50871441156775 | 14.357426070787312 | 0.7157353448275862 | estimate | latest published v1.5 row (v1.5.8 combined file) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): Frozen BASE_REFERENCES Qwen/Qwen3.8-27B market price (USD 0.42/M input, 3.00/M output) applied to 2,767,510 input tokens and zero output tokens over 1,624 decisions; self-hosted estimate, not measured hosted billing. Lineage verified by architecture; no pricing gate. The author method encodes each candidate independently. | 0.790 measured on-GPU, no network · 1.73 est. incl. network (2x + 150 ms) | https://huggingface.co/ZefanCai/Open-Jev-27B-v1.1 |
| outside caps | SimpleJev (Qwen3.5-0.8B, CPU) | simplejev-qwen3.5-0.8b | 31.59915405528791 | 13.000426205687345 | 50.197881904888476 | 58.929227255635794 | 70.31175519898045 | 0.009763559113300492 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 4.324 measured on-CPU, no network · 8.80 est. incl. network (2x + 150 ms) | https://github.com/featherless-ai/simple-jev |
| outside caps | Vega-U (Qwen3.5-4B reader) | vega-u-qwen3.5-4b | 44.31251291226081 | 8.801638441109716 | 79.8233873834119 | 60.651933418279626 | 64.90155987179207 | 0.014789313131313127 | estimate | ESTIMATE (*est.) through the frozen pricing_v15 path (module sha256 6202ef78...): vega-U is a 14.3 M decision engine on a FROZEN Qwen/Qwen3.5-4B reader, which resolves to deepinfra:Qwen/Qwen3.5-4B in the dated 25 Sep DeepInfra catalog snapshot (sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f), USD 0.03/M input and USD 0.15/M output. The model generates no text: 0 output tokens on every item, so cost = input tokens x USD 0.03/M. COMMON COST BASIS: published cost is recomputed over the same remaining 1,485 items as every other row; all 1,500 items count for Intelligence. ALL-ITEMS DIAGNOSTIC: USD 0.0367598 per 1,000 over all 1,500 items (0.569x cap). | 3.308 measured on-CPU, no network · 6.77 est. incl. network (2x + 150 ms) | https://huggingface.co/nandakishorm/vega-U |
| outside caps | Gutsy 0.8B v0.3 | gutsy-0.8b-v0.3 | 43.05068960177584 | 8.600978834309654 | 77.50040036924202 | 72.8877498600171 | 78.3542548810428 | 0.0052665333333333335 | estimate | ESTIMATE (*est.), input tokens only. The author's gutsy-inference 0.3.0 server emits no output-token field in any spelling; under the chief-of-staff decision of 9 Oct 2026 a missing output_tokens field on a native option-probability model that generates no text is treated as 0 output tokens, exactly as a server reporting a literal 0 is treated. Zero output is a verified property of this backend, not an assumption: it has no generation path at all (no generate/create_completion/n_predict/max_tokens/sampler); it prefills via llm.eval(tokens) and reads llama_get_logits_ith(ctx,-1) (r52 PREREG-AMENDMENT-4). Rate: the declared base Qwen/Qwen3.5-0.8B resolves in the frozen 25 Sep 2026 snapshot to deepinfra:Qwen/Qwen3.5-0.8B at USD 0.01 per 1M input and USD 0.05 per 1M output. This row's own measured usage on draw v1.6-regular-20261010-a2: 789,980 input tokens and 0 output over 1,500 decisions => USD 0.0052665333333333335 per 1,000 decisions. | 0.756 measured on-GPU, no network · 1.66 est. incl. network (2x + 150 ms) | https://huggingface.co/kouhxp/gutsy |
| outside caps | PurpleMIST-Mini-1.0 (Olaverse) | purplemist-mini-1.0 | 43.033098052325656 | 8.37974071175967 | 77.68645539289165 | 58.74735259439709 | 70.75879834655666 | 0.00943423569023569 | estimate | ESTIMATE (*est.), base-model market reference via the frozen module's Base-to-listed-model mapping (Qwen/Qwen3.5-2B-Base -> deepinfra:Qwen/Qwen3.5-2B), from the frozen 25 Sep 2026 snapshot through the UNMODIFIED pricing_v15 path (sha256 dae86740...): declared base Qwen/Qwen3.5-2B-Base -> deepinfra:Qwen/Qwen3.5-2B USD 0.02/M input, USD 0.1/M output. The author's server reports NO token usage, 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 (CoS deputy decision #18974; receipt r109/receipts/TOKCOUNT-PURPLEMIST-R109.json; independent recount PASS, r109/review-tokcount/REVIEW.json): 700,492 input tokens over the 1,485 answered items. Nothing is generated (one forward pass, pointer-head option probabilities), so output is 0 under the 9 Oct missing-output rule. Cost on the common 1,485-item basis. Estimated, not a 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 0.009434 per 1,000 decisions (0.146x the official cost cap USD 0.064595), against USD 0.009340 over all 1,500 items (v1.6.29). | 3.602 measured on-CPU, no network · 7.35 est. incl. network (2x + 150 ms) | https://huggingface.co/olaverse/PurpleMIST-Mini-1.0 |
| outside caps | Qwen3.5-0.8B Decision Model (Mourad Ghafiri) | mghafiri-qwen3.5-0.8b-decision-model | 42.62692401157305 | 7.297270616183145 | 77.95657740696296 | 55.90747410573378 | 79.5180630553356 | 0.00481649630541872 | estimate | v1.5.4 published cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): documented hosted-model estimate | 5.551 measured on-CPU, no network · 11.25 est. incl. network (2x + 150 ms) | https://huggingface.co/mghafiri/qwen3.5-0.8B-decision-model |
| outside caps | Vega-U (Qwen3.5-0.8B reader) | vega-u-qwen3.5-0.8b | 36.52005519443276 | 5.503418464051958 | 67.53669192481355 | 73.8348503635803 | 79.21519751338194 | 0.0049297710437710435 | estimate | ESTIMATE (*est.) through the frozen pricing_v15 path (module sha256 6202ef78...): vega-U is a 14.3 M decision engine on a FROZEN Qwen/Qwen3.5-0.8B reader, which resolves to deepinfra:Qwen/Qwen3.5-0.8B in the dated 25 Sep DeepInfra catalog snapshot (sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f), USD 0.01/M input and USD 0.05/M output - the same basis and the same numbers the live vegaml-0.8b row carries. The model generates no text: 0 output tokens on every item, so cost = input tokens x USD 0.01/M. 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 0.004930 per 1,000 decisions (0.076x the official cost cap USD 0.064595), against USD 0.012253 over all 1,500 items (v1.6.29). | 0.695 measured on-CPU, no network · 1.54 est. incl. network (2x + 150 ms) | https://huggingface.co/nandakishorm/vega-U |
| outside caps | Laya typed-decisions | laya-typed-decisions | 44.22061775531203 | 5.471270906587037 | 82.96996460403702 | 70.81965680870755 | 82.52942826196218 | 0.003822543103448276 | estimate | 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. | 0.864 measured on-CPU, no network · 1.88 est. incl. network (2x + 150 ms) | https://huggingface.co/convaiinnovations/laya/tree/1c5edc17a7acd8701df6fc341c0d179f1c62c982/typed-decisions |
| outside caps | open-jev-deberta-v3-large (local CPU) | open-jev-deberta-v3-large | 35.39766066495959 | 2.8293414417857914 | 67.96597988813339 | 67.62138193768119 | 77.5922354035203 | 0.005583743842364532 | estimate | 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 | 1.689 measured on-CPU, no network · 3.53 est. incl. network (2x + 150 ms) | https://github.com/kotoba-lang/typed-decisions |
| outside caps | modernbert-jev v2 | modernbert-jev-v2 | 34.64663304369169 | 2.6107469083924393 | 66.68251917899094 | 57.79924670989358 | 70.36992715953406 | 0.009720063333333334 | estimate | 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. | 2.600 measured on-CPU, no network · 5.35 est. incl. network (2x + 150 ms) | https://huggingface.co/ksonny4/modernbert-jev-v2 |
| outside caps | Laya (Convai Innovations, ModernBERT-large 421M) | laya | 32.50714989618597 | 1.7910341507793501 | 63.22326564159259 | 73.26302890816496 | 84.89402784895724 | 0.0031881034482758625 | estimate | 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 | 0.825 measured on-CPU, no network · 1.80 est. incl. network (2x + 150 ms) | https://huggingface.co/convaiinnovations/laya |
| outside caps | Mirror | mirror | 25.806386196486883 | 0 | 51.61277239297377 | 73.47211485993975 | 89.27300805786774 | 0.0022780665024630543 | estimate | 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 | 0.651 measured on-CPU, no network · 1.45 est. incl. network (2x + 150 ms) | https://github.com/Bluusun/Decision-API |

## API offerings (ranked separately on https://benchmarkheaven.com/jev-models/api)
Hosted endpoints we do not run ourselves. They are not ranked on the open-weights board; ranks are on the API leaderboard.
| Capability rank | Name | Key | Capability | Intelligence | Calibration | Speed | Cost axis | USD per 1,000 decisions | Basis kind | Basis | p50 latency (s) | Source URL |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| see API leaderboard | GPT-6 Luna (default medium reasoning effort) | gpt-6-luna | 98.44 | 96.88 | 100 | 71.77 | 38.41 | 0.1129575296108291 | tariff | 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) | 1.824 | https://benchmarkheaven.com/jev-models/gpt-6-luna |
| see API leaderboard | Qwen3.8 27B (Chutes TEE) | qwen3.8-27b | 98.19 | 96.38 | 100 | 57.43 | 0 | 2.1783887438423646 | estimate | 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) | 7.095 | https://benchmarkheaven.com/jev-models/qwen3.8-27b |
| see API leaderboard | GPT-6 Luna (low reasoning effort) | gpt-6-luna-low | 97.82 | 96.62 | 99.02 | 71.66 | 38.98 | 0.10814365482233503 | tariff | 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) | 1.854 | https://benchmarkheaven.com/jev-models/gpt-6-luna-low |
| see API leaderboard | DeepSeek V4.1 Flash (thinking default) | deepseek-flash | 97.35499999999999 | 94.71 | 100 | 68.07 | 19.56 | 0.480105076142132 | tariff | 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) | 2.038 | https://benchmarkheaven.com/jev-models/deepseek-flash |
| see API leaderboard | GPT-5.6 Luna (low reasoning effort) | gpt-5.6-luna | 94.49 | 94.07 | 94.91 | 73.12 | 30.16 | 0.2127451776649746 | tariff | 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) | 1.588 | https://benchmarkheaven.com/jev-models/gpt-5.6-luna |
| see API leaderboard | kettei-fat (Sqwish) | kettei-fat | 86.3099857359241 | 79.83918785475791 | 92.78078361709031 | 93.19124027169539 | 57.40773420072094 | 0.0262870707070707 | estimate | 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). | 0.20005768549162894 | https://docs.sqwish.ai |
| see API leaderboard | Drex Max (Nace AI) | drex-max | 84.22908268614222 | 78.89021516142793 | 89.56795021085651 | 89.80651018227358 | 55.336763284582396 | 0.030815878787878788 | listed | 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. | 0.2520033140317537 | https://drex.nace.ai |
| see API leaderboard | SubQ Snap (preview, SubQ) | subq-snap-preview | 81.47372650071756 | 72.25005247787018 | 90.69740052356492 | 86.40460211903752 | 57.693900899055016 | 0.025715993265993265 | estimate | 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 Florian on 14 Oct 2026 01:16 Berlin (board #19133). 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; CoS deputy decision #19174 rules 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: CoS deputy #18975 made the cost axis conditional on a separate frozen-pricing change passing re-freeze, unchanged-row proof and independent review; the chief of staff asked for it in #19133; 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). | 0.46557116050098557 | https://subq.ai |
| see API leaderboard | d1-ultra (Liquid AI) | liquid-d1-ultra | 80.47704277896817 | 69.70041168706942 | 91.25367387086692 | 88.99977436791615 | 62.300793859384434 | 0.01805677575757576 | estimate | 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). | 0.2313630650169216 | https://www.liquid.ai |
| see API leaderboard | Bespoke Nimble-v4-max (Bespoke Labs) | nimble-v4-max | 79.71745310348602 | 68.18491883484498 | 91.24998737212707 | 91.45091520080334 | 60.74282184772803 | 0.020350383838383838 | estimate | 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). | 0.254514531508903 | https://bespokelabs.ai |
| see API leaderboard | Amos 1 large (Scira) | amos-1-large | 79.35913670315466 | 69.72341288828827 | 88.99486051802106 | 83.61644312651299 | 30.04968825138303 | 0.21462339393939392 | estimate | 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). | 0.5090118609950878 | https://huggingface.co/mukaddamzaid/amos-1-large |
| see API leaderboard | wity-1 (Wity, reasoning auto) | wity-1 | 79.11559492280433 | 70.28143483604077 | 87.9497550095679 | 71.79000307789984 | 58.8349720563948 | 0.023559582938388626 | tariff | 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 | 1.5746641159057617 | https://wity.alphanimble.com/ |
| see API leaderboard | Myles-G 31B (Launchables) | launchables-myles-g-31b | 78.81859215262476 | 73.64186833978687 | 83.99531596546265 | 86.60775159362743 | 51.11161516579885 | 0.0426199393939394 | estimate | 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). | 0.4372560594929382 | https://www.launchables.ai/licensing |
| see API leaderboard | SPX-CD Pro (SurdAI, hosted /v1/systemone, Oct-4 checkpoint) | spx-cd-pro | 78.63514339971708 | 64.85693822910903 | 92.41334857032513 | 77.96292698098645 | 43.076275488494026 | 0.07896915368991199 | estimate | 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 | 1.1281277425587177 | https://surd.ai |
| see API leaderboard | Sage 1.3.0 (Levanto Labs) | sage-1.3.0 | 78.59554275566781 | 65.59290944525489 | 91.59817606608073 | 93.61638181239104 | 58.231203787388296 | 0.024677048070413 | estimate | 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) | 0.15245595201849937 | https://levanto.ai |
| see API leaderboard | Fastino GLiDE | fastino-glide | 77.025 | 66.1 | 87.95 | 77.48 | 35.86 | 0.13735482233502538 | tariff | 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) | 0.743 | not published |
| see API leaderboard | Jev 1.14.0 (TypeSafe AI) | jev-1.14.0 | 75.01246945698011 | 59.942981848430136 | 90.08195706553009 | 90.96457246082724 | 54.815314343766296 | 0.032074226262626265 | estimate | 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). | 0.2238923824916128 | https://docs.typesafe.ai/models |
| see API leaderboard | Sage 1.4 (Levanto Labs) | sage-1.4.0 | 74.77338506693098 | 65.88206323551285 | 83.66470689834914 | 68.80003106294234 | 59.47868179714396 | 0.022423872053872054 | estimate | 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 board #18983 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). | 1.7410984199959785 | https://levanto.ai/ |
| see API leaderboard | d1 (Liquid AI) | liquid-d1 | 74.4216922125323 | 62.05738930369758 | 86.78599512136702 | 89.22497122158413 | 62.84038569943876 | 0.017324224779959378 | tariff | 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 | 0.26846902817487717 | https://www.liquid.ai/blog/d1-decision-model |
| see API leaderboard | SPX-CD Flash (SurdAI, hosted /v1/systemone, Oct-4 checkpoint) | spx-cd-flash | 74.31544056451752 | 58.43051843707937 | 90.20036269195566 | 79.39117060033084 | 52.107175358413464 | 0.039484576844955994 | estimate | 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 | 0.9804960675537586 | https://surd.ai |
| see API leaderboard | Mercury Decide (Inception; System One decisions API, served free on OpenRouter as inception/mercury-decide:free) | mercury-decide | 73.62521498320774 | 65.90701345473542 | 81.34341651168008 | 85.47752366478576 | 62.08313746631354 | 0.018360961408259987 | estimate | 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. RELEASE-LANE LINE, recorded and not acted on, and it is RUN 57'S SHAPE EXACTLY - one mapping line, not a new price: 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. | 0.33213604614138603 | https://openrouter.ai/inception/mercury-decide:free |
| see API leaderboard | Celeris-1 Decision (Celeris) | celeris-1-decision | 73.62245264836737 | 58.890192862872325 | 88.35471243386242 | 86.97194897947064 | 60.373088134904265 | 0.020936161616161617 | estimate | 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). | 0.4326209495193325 | https://docs.celeris.ai |
| see API leaderboard | OpenAI Decisions (gpt-6-luna) | openai-decisions | 73.5 | 56.9 | 90.1 | 89.25 | 48.6 | 0.0517 | tariff | 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. | 0.298 | not published |
| see API leaderboard | kettei-chub (Sqwish) | kettei-chub | 73.14290542941187 | 62.182714942969874 | 84.10309591585386 | 95.09167642607092 | 62.69047197239138 | 0.017524713804713803 | estimate | 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). | 0.16667446299106814 | https://docs.sqwish.ai |
| see API leaderboard | Autoloops – Gemma 4 31B IT | kushal-gemma4-31b-it-autoloops | 72.705 | 58.1 | 87.31 | 84.51 | 40.47 | 0.09648663282571913 | tariff | 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) | 0.584 | https://benchmarkheaven.com/jev-models/kushal-gemma4-31b-it-autoloops |
| see API leaderboard | SimpleJev Qwen3.8-27B | simplejev-qwen3.8-27b | 71.605 | 52.88 | 90.33 | 76.93 | 14.9 | 0.686803448275862 | estimate | 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) | 1.212 | https://github.com/featherless-ai/simple-jev |
| see API leaderboard | Microsoft-Decision-1 (Azure Foundry) | microsoft-decision-1 | 70.78415790879176 | 57.16958510591244 | 84.39873071167106 | 81.54517543213291 | 61.452075716621465 | 0.0192721784989858 | tariff | Advertised Azure Foundry tariff .042/M input, output free; genuine native usage, billed invoice unknown | 0.45925578847527504 | https://commandline.microsoft.com/microsoft-decision-1-model-foundry/ |
| see API leaderboard | Myles-G 12B (Launchables) | launchables-myles-g-12b | 70.65140999372262 | 57.25409127121432 | 84.04872871623093 | 85.64932567577961 | 57.35738348807602 | 0.02638885521885522 | estimate | ESTIMATE, nearest-listed-sibling reference: declared base google/gemma-4-12B-it is UNLISTED in frozen pricing_v15 (sha256 a8a05890ddab2d39d1a1bd543f2b5c83842bdc70aa54766788bc89b7a8a448d7); no exact base reference exists. Same-base published precedents arcbright-elf-1-12b, chon-1.0-12b and metask_jev_rain_12b_trained use google/gemma-3-12b-it, USD 0.05/M input and 0.15/M output. This row uses that same sibling ESTIMATE, not a vendor bill or quote. Launchables list price GBP 0.0015 per million input tokens, output free, is shown for information and is not used for the Cost axis. Applied to provider-reported usage. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 779,294 input / 1,485 output tokens. All-item usage 989,607 input / 1,500 output over 1,500 items. | 0.479910550522618 | https://www.launchables.ai/licensing |
| see API leaderboard | Instinct (ZooWork, Qwen3.8-27B) | instinct | 70.47 | 47.77 | 93.17 | 87.2 | 64.41 | 0.015355329949238578 | tariff | 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) | 0.292 | https://github.com/fstandhartinger/jevbench/issues/69 |
| see API leaderboard | attest-31B-v1.0 | attest-31b-v1.0 | 69.54321289428123 | 65.23849472031306 | 73.84793106824938 | 97.10993413996468 | 50.72941039177032 | 0.04388872727272727 | estimate | 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). | 0.10787388649987406 | https://attest.engineering/model |
| see API leaderboard | SimpleJev Qwen3.6-35B-A3B | simplejev-qwen3.6-35b-a3b | 68.295 | 56.29 | 80.3 | 76.7 | 35.15 | 0.1451164454277286 | estimate | 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) | 1.25 | https://github.com/featherless-ai/simple-jev |
| see API leaderboard | HS1 0.1.2 (Hyland) | hyland-hs1 | 65.03939854649335 | 53.57948799459325 | 76.49930909839343 | 93.74328417303443 | 49.09977911142268 | 0.04973636363636364 | estimate | 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). | 0.19411396997747943 | https://www.hyland.com |
| see API leaderboard | Bespoke Nimble-v4 (Bespoke Labs) | nimble-v4 | 64.01968095926979 | 41.405096245385266 | 86.63426567315432 | 88.39856712562789 | 60.74282184772803 | 0.020350383838383838 | estimate | 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 CoS deputy decision #19084 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). | 0.23237984799197875 | https://bespokelabs.ai |
| see API leaderboard | Vansa-3.4 (Vansa, hosted System One API) | vansa-3.4 | 63.685 | 40.85 | 86.52 | 95.49 | 61.92 | 0.018590751269035533 | tariff | 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) | 0.15 | https://vansa.org |
| see API leaderboard | Amos 1 medium (Scira) | amos-1-medium | 63.175111580896406 | 38.36362733252365 | 87.98659582926916 | 86.48270625779318 | 48.74716696332005 | 0.05110080808080808 | estimate | 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). | 0.44348932350840187 | https://huggingface.co/mukaddamzaid/amos-1-medium |
| see API leaderboard | Gemini 3.1 Flash-Lite | gemini-3.1-flash-lite | 59.75 | 58.57 | 60.93 | 79.03 | 29.17 | 0.22953722504230117 | tariff | 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) | 0.957 | https://benchmarkheaven.com/jev-models/gemini-3.1-flash-lite |
| see API leaderboard | openjev-sglang (Qwen3.6-35B-A3B on SGLang) | openjev-sglang | 58.67 | 38.77 | 78.57 | 77.71 | 35.63 | 0.13981028325123152 | estimate | 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) | 1.265 | https://github.com/ekzhang/openjev-sglang |
| see API leaderboard | Instinct Dual 4B | instinct-dual-4b | 57.21 | 28.34 | 86.08 | 88.6 | 75.03 | 0.006799306260575296 | tariff | 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) | 0.231 | https://github.com/fstandhartinger/jevbench/issues/123 |
| see API leaderboard | system-one-open (Gemma 4 E2B LoRA on an L4) | system-one-open | 51.94 | 27.94 | 75.94 | 79.32 | 68.33 | 0.011367487684729064 | estimate | 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 | 1.02 | https://github.com/mithalouni/system-one-open |
| see API leaderboard | decision-machine-1 (milliseconds.ai) | decision-machine-1 | 46.13 | 13.11 | 79.15 | 92.97 | 58.33 | 0.02449299492385787 | tariff | 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) | 0.167 | https://www.milliseconds.ai |
| see API leaderboard | Fastino GLiNER-2.5-Decide (hosted API) | fastino-gliner-2-5-decide | 34.57211625543646 | 7.011351364482595 | 62.13288114639033 | 84.7052606179839 | 45.81558630529788 | 0.06399515910629654 | mixed | 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 | 0.4409300498664379 | https://fastino.ai/ |
| see API leaderboard | JevAct (einptein, jev1-2b-v2) | jevact | 28.77 | 10.1 | 47.44 | 76.45 | 68.56 | 0.011167487684729064 | estimate | 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 | 0.801 | https://github.com/fstandhartinger/jevbench/issues/66 |

## Guides and comparisons
- https://benchmarkheaven.com/jev-models/alternatives
- https://benchmarkheaven.com/jev-models/how-to-choose
- https://benchmarkheaven.com/jev-models/open-source-jev
- [Jev vs Imajev-4B](https://benchmarkheaven.com/jev-models/jev-vs-imajev)
- [Jev vs Plumb-4B](https://benchmarkheaven.com/jev-models/jev-vs-plumb)
- [Jev vs decider-4b v2](https://benchmarkheaven.com/jev-models/jev-vs-decider-4b-v2)
- [Jev vs JevK5](https://benchmarkheaven.com/jev-models/jev-vs-jevk5)
- [Jev vs Cygnet](https://benchmarkheaven.com/jev-models/jev-vs-cygnet)
- [Jev vs Hopper](https://benchmarkheaven.com/jev-models/jev-vs-hopper)
- [Jev vs Winnow-12B Q8](https://benchmarkheaven.com/jev-models/jev-vs-winnow-12b-q8)
- [Jev vs reflex 4B](https://benchmarkheaven.com/jev-models/jev-vs-reflex-4b)
- [Jev vs Laya](https://benchmarkheaven.com/jev-models/jev-vs-laya)
- [Jev vs Xor 26B-A4B](https://benchmarkheaven.com/jev-models/jev-vs-xor-26b-a4b)
- [Jev vs Surogate Rune 26B-A4B v3](https://benchmarkheaven.com/jev-models/jev-vs-surogate-rune-26b-a4b-v3-v16)
- [Jev vs Torchcast Decision 12B](https://benchmarkheaven.com/jev-models/jev-vs-torchcast-decision-12b)
- [Jev vs Jev-Omni](https://benchmarkheaven.com/jev-models/jev-vs-jev-omni)
- [Jev vs Quyet-1.0-Large](https://benchmarkheaven.com/jev-models/jev-vs-quyet-1-0-large)
- [Jev vs decisio v0.8.0 on gemma-4-31B-it](https://benchmarkheaven.com/jev-models/jev-vs-decisio-gemma-4-31b-v080)
- [Jev vs Sage 1.3.0](https://benchmarkheaven.com/jev-models/jev-vs-sage-1.3.0)
- [Jev vs deck-31B](https://benchmarkheaven.com/jev-models/jev-vs-deck31b)
- [Jev vs René-1 31B FP8](https://benchmarkheaven.com/jev-models/jev-vs-rene-1-31b-fp8)
- [Jev vs H2O-Lightning-4B v1.1](https://benchmarkheaven.com/jev-models/jev-vs-h2o-lightning-4b)
- [Jev vs SPX-CD Flash](https://benchmarkheaven.com/jev-models/jev-vs-spx-cd-flash)
- [Jev vs kettei-fat](https://benchmarkheaven.com/jev-models/jev-vs-kettei-fat)
- [Jev vs Chon-1.0-12B](https://benchmarkheaven.com/jev-models/jev-vs-chon-1.0-12b)
- [Jev vs H2O-Lightning-31B v1.0.7](https://benchmarkheaven.com/jev-models/jev-vs-h2o-lightning-31b)
- [Jev vs Drex Max](https://benchmarkheaven.com/jev-models/jev-vs-drex-max)
- [Jev vs basal-omni-1.0-ultra](https://benchmarkheaven.com/jev-models/jev-vs-basal-omni-1.0-ultra)
- [Jev vs Jeff-1.0-Large](https://benchmarkheaven.com/jev-models/jev-vs-jeff_1_0_large)
- [Jev vs SubQ Snap](https://benchmarkheaven.com/jev-models/jev-vs-subq-snap-preview)
## Model pages
- https://benchmarkheaven.com/jev-models/sage-1.3.0
- https://benchmarkheaven.com/jev-models/mercury-decide
- https://benchmarkheaven.com/jev-models/decisio-gemma-4-31b-v080
- https://benchmarkheaven.com/jev-models/jev-1.13.0
- https://benchmarkheaven.com/jev-models/quyet-1-0-large
- https://benchmarkheaven.com/jev-models/decider-12b
- https://benchmarkheaven.com/jev-models/wity-1
- https://benchmarkheaven.com/jev-models/decider-12b-v1
- https://benchmarkheaven.com/jev-models/torchcast-decision-12b
- https://benchmarkheaven.com/jev-models/winnow-12b
- https://benchmarkheaven.com/jev-models/deck31b
- https://benchmarkheaven.com/jev-models/cygnet
- https://benchmarkheaven.com/jev-models/decisio-gemma-4-12b-v080
- https://benchmarkheaven.com/jev-models/decisio-gemma-4-12b-v090
- https://benchmarkheaven.com/jev-models/jev-omni
- https://benchmarkheaven.com/jev-models/xor-26b-a4b
- https://benchmarkheaven.com/jev-models/bobcat-flash-1.2
- https://benchmarkheaven.com/jev-models/spx-cd-flash
- https://benchmarkheaven.com/jev-models/decider-chat-gemma4-31b
- https://benchmarkheaven.com/jev-models/hopper-12b-trained
- https://benchmarkheaven.com/jev-models/surogate-rune-26b-a4b-v3-v16
- https://benchmarkheaven.com/jev-models/gev-26b-decide
- https://benchmarkheaven.com/jev-models/spx-cd-omni
- https://benchmarkheaven.com/jev-models/diffusion-jev
- https://benchmarkheaven.com/jev-models/rene-1-31b-fp8
- https://benchmarkheaven.com/jev-models/aplomb-1
- https://benchmarkheaven.com/jev-models/bespoke-nimble-9b-v3
- https://benchmarkheaven.com/jev-models/apus-openjev-v1-9b
- https://benchmarkheaven.com/jev-models/plumb-4b
- https://benchmarkheaven.com/jev-models/spx-cd-pro
- https://benchmarkheaven.com/jev-models/messier-one-v0.2
- https://benchmarkheaven.com/jev-models/spark-s1-4b-v6
- https://benchmarkheaven.com/jev-models/swanone
- https://benchmarkheaven.com/jev-models/quyet-1-0-medium
- https://benchmarkheaven.com/jev-models/lev
- https://benchmarkheaven.com/jev-models/ninfer-qwen3.8-flash-next
- https://benchmarkheaven.com/jev-models/jev-local
- https://benchmarkheaven.com/jev-models/decider-4b-v2
- https://benchmarkheaven.com/jev-models/classone-qwen3.5-9b
- https://benchmarkheaven.com/jev-models/jevk5-v0.3-4b
- https://benchmarkheaven.com/jev-models/metask-jev-4b
- https://benchmarkheaven.com/jev-models/deck-4b-v1-0
- https://benchmarkheaven.com/jev-models/clef-flash
- https://benchmarkheaven.com/jev-models/evalengine-decision-4b
- https://benchmarkheaven.com/jev-models/janus-4b
- https://benchmarkheaven.com/jev-models/qwen35-9b-jev-data-mix-v2
- https://benchmarkheaven.com/jev-models/jevk5-v02
- https://benchmarkheaven.com/jev-models/jevone
- https://benchmarkheaven.com/jev-models/typecastlm-1.4.0
- https://benchmarkheaven.com/jev-models/hopper
- https://benchmarkheaven.com/jev-models/imajev-4b-rtx5090-a2
- https://benchmarkheaven.com/jev-models/malkuth-4b
- https://benchmarkheaven.com/jev-models/jqv
- https://benchmarkheaven.com/jev-models/jpt-35b-a3b
- https://benchmarkheaven.com/jev-models/decider-35b-a3b
- https://benchmarkheaven.com/jev-models/coco-decision-4b
- https://benchmarkheaven.com/jev-models/manchego-v21-v15
- https://benchmarkheaven.com/jev-models/decision-4b-v12
- https://benchmarkheaven.com/jev-models/raw-qwen3-4b-instruct-2507
- https://benchmarkheaven.com/jev-models/jev-27b-autotrust
- https://benchmarkheaven.com/jev-models/torchcast-decision-27b
- https://benchmarkheaven.com/jev-models/nimble-9b
- https://benchmarkheaven.com/jev-models/reflex-4b
- https://benchmarkheaven.com/jev-models/pplx-decider-v1-1-27b
- https://benchmarkheaven.com/jev-models/jade
- https://benchmarkheaven.com/jev-models/typecastlm
- https://benchmarkheaven.com/jev-models/jebadiah-27b
- https://benchmarkheaven.com/jev-models/decision-2-0-vega-27b
- https://benchmarkheaven.com/jev-models/decision-4b-v11
- https://benchmarkheaven.com/jev-models/autojev-27b
- https://benchmarkheaven.com/jev-models/eikos-27b
- https://benchmarkheaven.com/jev-models/ninfer-qwen3.8-27b
- https://benchmarkheaven.com/jev-models/openjev-thinking
- https://benchmarkheaven.com/jev-models/open-jev-zefan-9b
- https://benchmarkheaven.com/jev-models/standardone-8b
- https://benchmarkheaven.com/jev-models/clef
- https://benchmarkheaven.com/jev-models/jev-qwen3.5-9b-base-nvfp4
- https://benchmarkheaven.com/jev-models/semif-qwen3.5-4b
- https://benchmarkheaven.com/jev-models/bev-bonsai-27b
- https://benchmarkheaven.com/jev-models/litjev
- https://benchmarkheaven.com/jev-models/localjev-qwen3.5-4b
- https://benchmarkheaven.com/jev-models/system-one-sg
- https://benchmarkheaven.com/jev-models/kev-8b
- https://benchmarkheaven.com/jev-models/decision-2b
- https://benchmarkheaven.com/jev-models/opensourcejev-qwen35-4b-q4km
- https://benchmarkheaven.com/jev-models/open-alternative-jev
- https://benchmarkheaven.com/jev-models/raw-qwen3-8b
- https://benchmarkheaven.com/jev-models/decider-2b
- https://benchmarkheaven.com/jev-models/nemotron-diffusion-8b
- https://benchmarkheaven.com/jev-models/kev-4b
- https://benchmarkheaven.com/jev-models/kev-27b
- https://benchmarkheaven.com/jev-models/malkuth-2b
- https://benchmarkheaven.com/jev-models/qwen3-reranker-4b
- https://benchmarkheaven.com/jev-models/simplejev-qwen3.8-27b-selfhost
- https://benchmarkheaven.com/jev-models/zerank-2
- https://benchmarkheaven.com/jev-models/exaone-4.0-1.2b-jev-v0.3
- https://benchmarkheaven.com/jev-models/open-jev-zefan-2b
- https://benchmarkheaven.com/jev-models/open-jev-zefan-27b-v1.1
- https://benchmarkheaven.com/jev-models/raw-phi-4-mini
- https://benchmarkheaven.com/jev-models/simplejev-qwen3.5-0.8b
- https://benchmarkheaven.com/jev-models/smalljev
- https://benchmarkheaven.com/jev-models/kev-0.6b
- https://benchmarkheaven.com/jev-models/jul-fast
- https://benchmarkheaven.com/jev-models/opendecision
- https://benchmarkheaven.com/jev-models/classone-gemma4-e2b
- https://benchmarkheaven.com/jev-models/mghafiri-qwen3.5-0.8b-decision-model
- https://benchmarkheaven.com/jev-models/fastino-gliner-2-5-decide
- https://benchmarkheaven.com/jev-models/deem-0.8-v1
- https://benchmarkheaven.com/jev-models/decision-fast
- https://benchmarkheaven.com/jev-models/quyet-1-0-small-en
- https://benchmarkheaven.com/jev-models/laya-typed-decisions
- https://benchmarkheaven.com/jev-models/watersheep
- https://benchmarkheaven.com/jev-models/openjev-verdict-1.4
- https://benchmarkheaven.com/jev-models/raw-qwen3-0.6b
- https://benchmarkheaven.com/jev-models/lev-350m
- https://benchmarkheaven.com/jev-models/raw-qwen3-1.7b
- https://benchmarkheaven.com/jev-models/kev-0.5b
- https://benchmarkheaven.com/jev-models/openjev-verdict
- https://benchmarkheaven.com/jev-models/quyet-1-0-small
- https://benchmarkheaven.com/jev-models/watt-flash-0.1
- https://benchmarkheaven.com/jev-models/quyet-1-0-tiny
- https://benchmarkheaven.com/jev-models/open-jev-deberta-v3-large
- https://benchmarkheaven.com/jev-models/verdict-small
- https://benchmarkheaven.com/jev-models/tacet-sonata
- https://benchmarkheaven.com/jev-models/laya
- https://benchmarkheaven.com/jev-models/mxbai-rerank-base-v2
- https://benchmarkheaven.com/jev-models/laya-multilingual
- https://benchmarkheaven.com/jev-models/certo
- https://benchmarkheaven.com/jev-models/clm-8b
- https://benchmarkheaven.com/jev-models/bge-reranker-v2-m3
- https://benchmarkheaven.com/jev-models/gte-reranker-modernbert-base
- https://benchmarkheaven.com/jev-models/mirror
- https://benchmarkheaven.com/jev-models/open-jev-json-canvas-joshuasp
- https://benchmarkheaven.com/jev-models/apus-openjev-v1-35b-a3b
- https://benchmarkheaven.com/jev-models/bb-qwen3.5-4b-lora
- https://benchmarkheaven.com/jev-models/wity-1-always
- https://benchmarkheaven.com/jev-models/wity-1-off
- https://benchmarkheaven.com/jev-models/classifier-dev-fast
- https://benchmarkheaven.com/jev-models/instinct
- https://benchmarkheaven.com/jev-models/vansa-3.4
- https://benchmarkheaven.com/jev-models/gpt-6-luna-low
- https://benchmarkheaven.com/jev-models/autoloops-gemma-4-31b-it
- https://benchmarkheaven.com/jev-models/gpt-6-luna
- https://benchmarkheaven.com/jev-models/fastino-glide
- https://benchmarkheaven.com/jev-models/simplejev-qwen3.6-35b-a3b
- https://benchmarkheaven.com/jev-models/gpt-5.6-luna
- https://benchmarkheaven.com/jev-models/instinct-dual-4b
- https://benchmarkheaven.com/jev-models/gemini-3.1-flash-lite
- https://benchmarkheaven.com/jev-models/system-one-open
- https://benchmarkheaven.com/jev-models/openjev-sglang
- https://benchmarkheaven.com/jev-models/deepseek-flash
- https://benchmarkheaven.com/jev-models/simplejev-qwen3.8-27b
- https://benchmarkheaven.com/jev-models/decision-machine-1
- https://benchmarkheaven.com/jev-models/jevact
- https://benchmarkheaven.com/jev-models/qwen3.8-27b
- https://benchmarkheaven.com/jev-models/openai-decisions
- https://benchmarkheaven.com/jev-models/liquid-d1
- https://benchmarkheaven.com/jev-models/microsoft-decision-1
- https://benchmarkheaven.com/jev-models/decisor_4b
- https://benchmarkheaven.com/jev-models/jeff_1_0_large
- https://benchmarkheaven.com/jev-models/metask-jev-rain-4b
- https://benchmarkheaven.com/jev-models/metask_jev_rain_12b
- https://benchmarkheaven.com/jev-models/ryotide_qwen9
- https://benchmarkheaven.com/jev-models/wald-4b-v2
- https://benchmarkheaven.com/jev-models/blink-v0-3-26b-a4b-nvfp4
- https://benchmarkheaven.com/jev-models/classone-gemma4-e4b
- https://benchmarkheaven.com/jev-models/classone-qwen3.5-2b
- https://benchmarkheaven.com/jev-models/classone-qwen3.5-4b
- https://benchmarkheaven.com/jev-models/clef-omni
- https://benchmarkheaven.com/jev-models/gutsy-0.8b-v0.3
- https://benchmarkheaven.com/jev-models/jiwo-0.8b
- https://benchmarkheaven.com/jev-models/jiwo-4b
- https://benchmarkheaven.com/jev-models/liquid-d1-3b
- https://benchmarkheaven.com/jev-models/liquid-d1-omni-600m
- https://benchmarkheaven.com/jev-models/modernbert-jev-v2
- https://benchmarkheaven.com/jev-models/rsi-jev-v6.1-vl-27b
- https://benchmarkheaven.com/jev-models/standard_one_8b_sh
- https://benchmarkheaven.com/jev-models/vegaml-0.8b
- https://benchmarkheaven.com/jev-models/vegaml-4b
- https://benchmarkheaven.com/jev-models/wald-4b-v2-1
- https://benchmarkheaven.com/jev-models/amos-1-large
- https://benchmarkheaven.com/jev-models/amos-1-medium
- https://benchmarkheaven.com/jev-models/arcbright-elf-1-12b
- https://benchmarkheaven.com/jev-models/arsenelupin-v1.1
- https://benchmarkheaven.com/jev-models/attest-31b-v1.0
- https://benchmarkheaven.com/jev-models/basal-omni-1.0-ultra
- https://benchmarkheaven.com/jev-models/blink-4b
- https://benchmarkheaven.com/jev-models/bongard-mini
- https://benchmarkheaven.com/jev-models/celeris-1-decision
- https://benchmarkheaven.com/jev-models/chon-1.0-12b
- https://benchmarkheaven.com/jev-models/chon-1.0-e4b
- https://benchmarkheaven.com/jev-models/compass-0.2.0
- https://benchmarkheaven.com/jev-models/decision2-lux-9b
- https://benchmarkheaven.com/jev-models/drex-max
- https://benchmarkheaven.com/jev-models/goddragon-arbiter-4b
- https://benchmarkheaven.com/jev-models/h2o-lightning-31b
- https://benchmarkheaven.com/jev-models/h2o-lightning-4b
- https://benchmarkheaven.com/jev-models/hyland-hs1
- https://benchmarkheaven.com/jev-models/imajev-2b
- https://benchmarkheaven.com/jev-models/imajev-9b
- https://benchmarkheaven.com/jev-models/iota-decide-4b
- https://benchmarkheaven.com/jev-models/jeb-35b-a3b
- https://benchmarkheaven.com/jev-models/jebadiah-9b-v2
- https://benchmarkheaven.com/jev-models/jebadiah-9b-v2-1
- https://benchmarkheaven.com/jev-models/jev-1.14.0
- https://benchmarkheaven.com/jev-models/jevora-9b-v1
- https://benchmarkheaven.com/jev-models/jpt-4b
- https://benchmarkheaven.com/jev-models/julia-1
- https://benchmarkheaven.com/jev-models/kahn1
- https://benchmarkheaven.com/jev-models/kettei-chub
- https://benchmarkheaven.com/jev-models/kettei-fat
- https://benchmarkheaven.com/jev-models/kev-9b-v2
- https://benchmarkheaven.com/jev-models/krino-2b
- https://benchmarkheaven.com/jev-models/launchables-myles-g-12b
- https://benchmarkheaven.com/jev-models/launchables-myles-g-31b
- https://benchmarkheaven.com/jev-models/liquid-d1-ultra
- https://benchmarkheaven.com/jev-models/messier-one-v0.4
- https://benchmarkheaven.com/jev-models/metask_jev_rain_12b_trained
- https://benchmarkheaven.com/jev-models/metask_jev_rain_31b_trained
- https://benchmarkheaven.com/jev-models/metask_jev_rain_31b_trained_v3
- https://benchmarkheaven.com/jev-models/microsoft-decision-1-a2-20261012
- https://benchmarkheaven.com/jev-models/nagi-enormous
- https://benchmarkheaven.com/jev-models/nagi-huge
- https://benchmarkheaven.com/jev-models/nemotron-deciffusion-8b-v2
- https://benchmarkheaven.com/jev-models/nev-2b
- https://benchmarkheaven.com/jev-models/nimble-v4
- https://benchmarkheaven.com/jev-models/nimble-v4-max
- https://benchmarkheaven.com/jev-models/noma
- https://benchmarkheaven.com/jev-models/phocinae-largha-150m-v1
- https://benchmarkheaven.com/jev-models/pjev-2b
- https://benchmarkheaven.com/jev-models/pplx-decider-flash-v1.2
- https://benchmarkheaven.com/jev-models/pplx-decider-v1.2
- https://benchmarkheaven.com/jev-models/purplemist-flash-1.0
- https://benchmarkheaven.com/jev-models/purplemist-mini-1.0
- https://benchmarkheaven.com/jev-models/reflex-instinct-0.6b
- https://benchmarkheaven.com/jev-models/respan-span-02
- https://benchmarkheaven.com/jev-models/roudone-decider-4b
- https://benchmarkheaven.com/jev-models/ryotide-gemma
- https://benchmarkheaven.com/jev-models/ryotide-qwen
- https://benchmarkheaven.com/jev-models/sage-1.4.0
- https://benchmarkheaven.com/jev-models/seb-9b
- https://benchmarkheaven.com/jev-models/shisa-de-1
- https://benchmarkheaven.com/jev-models/standard_one_3b_sh
- https://benchmarkheaven.com/jev-models/startlux-decision-v1.1-27b-autothink
- https://benchmarkheaven.com/jev-models/startlux-decision-v1.1-27b-instant
- https://benchmarkheaven.com/jev-models/subq-snap-preview
- https://benchmarkheaven.com/jev-models/tde-general-large
- https://benchmarkheaven.com/jev-models/torchcast-decision-4b-v1.1
- https://benchmarkheaven.com/jev-models/torchcast-decision-4b-v1.2
- https://benchmarkheaven.com/jev-models/trio-spark-v1.3
- https://benchmarkheaven.com/jev-models/uniphore-4b-decider
- https://benchmarkheaven.com/jev-models/vega-u-qwen3.5-0.8b
- https://benchmarkheaven.com/jev-models/vega-u-qwen3.5-4b
- https://benchmarkheaven.com/jev-models/verdin-4b
- https://benchmarkheaven.com/jev-models/vllm-sr-d3
- https://benchmarkheaven.com/jev-models/vllm-sr-d3-turbo
- https://benchmarkheaven.com/jev-models/autojev-27b-rtxpro6000-a2
- https://benchmarkheaven.com/jev-models/surogate-rune-26b-a4b-v3-rtxpro6000-a2
- https://benchmarkheaven.com/jev-models/reflex-27b
- https://benchmarkheaven.com/jev-models/ninfer-qwen3.8-27b-t1.5
- https://benchmarkheaven.com/jev-models/jeff
- https://benchmarkheaven.com/jev-models/von-395m
- https://benchmarkheaven.com/jev-models/bosun-v31-0.6b
- https://benchmarkheaven.com/jev-models/gliner2.5-multi
- https://benchmarkheaven.com/jev-models/gliner2.5-small
- https://benchmarkheaven.com/jev-models/gliner2-large
- https://benchmarkheaven.com/jev-models/gliner2
- https://benchmarkheaven.com/jev-models/needle-3
- https://benchmarkheaven.com/jev-models/needle-3-tools
- https://benchmarkheaven.com/jev-models/imajev_4b
- https://benchmarkheaven.com/jev-models/jobe-qwen3.5-4b
- https://benchmarkheaven.com/jev-models/openjev-razorback16

## API
- https://benchmarkheaven.com/api/jevbench/latest
- https://benchmarkheaven.com/api/jevbench/v1.6.39
- https://benchmarkheaven.com/llms.txt