Super Deciderv1.7 · 10 s rule

Composite (secondary)

78.5 · Not ranked on this board

Cost and measurement conditions

$0.016 per 1,000 decisions (not published)

Launch list price: the launch list price stated on the vendor's public product page (singularitycompute.com/super-decider, fetched 22 Oct 2026: USD 0.03 per 1M input tokens, output tokens free); the ASI:Cloud catalogue itself was not reachable to us. The page lists one price for the product and no separate rate for thinking mode; we apply it to both rows. Counted as an operator launch list price by the frozen pricing module pricing_v15 (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7, re-frozen on 22 Oct 2026 with launch-list keys added and nothing else changed; the launch-list line and the rates for this product are identical in the earlier module 81e3578f), which carries a launch-list-price line for this product. Applied to the tokens reported by the provider's API; not a base-model estimate. Over all 1,500 items the frozen resolve() gives USD 0.039197 per 1,000 decisions; the official cost axis uses the common 1,485-item basis, where it is USD 0.015531 per 1,000 decisions. COMMON COST BASIS (v1.6.30): common 1,485 items, excluding the 15 longest for every row; 768,771 input / 322,546 output tokens. All-item usage 1,959,826 input / 328,348 output over 1,500 items.

p50 latency: 0.2 s

none (API)

Paid fast-lane evaluation, published at the provider's request (the order was private and was made public at the customer's request). Vendor-hosted, first-party Singularity Compute API (llm.a.singularitycompute.com, POST /v1/systemone, model `super-decider`), through the unchanged TypeSafe adapter; latency measured from Sandy (Germany) without API adjustment. Full S1200+P300 draw of v1.6-regular-20261012 through the unchanged run_api_full_v161.py; the full-set exposure gate logged all 1,200 sealed items to provider `singularitycompute` (no proxy) before the first request. All 1,500 of 1,500 requests answered with HTTP 200 and the served model was `super-decider` on every row; native probabilities on all three question types; one serial pass, no retries. Run with the provider's own API key; hosting implementation and deployed weights were not independently inspected.

Model and serving disclosure

Vendor-hosted API: no vendor code ran locally. Model, as pinned by the provider: Singularity-Compute/super-decider at 7f8b3f76e88aba6aa1159da75eb577ea990d2af1 on Hugging Face (public, not gated), a discrete-diffusion decision model fine-tuned from google/diffusiongemma-26B-A4B-it and served from an NVFP4 checkpoint (card statement). Serving code, as pinned by the provider: github.com/Singularity-Compute/super-decider at 10bf7eddcbd3b5f6dddf27e868231b2e1140dfba (pinned vLLM nightly with the provider's patches and a structured server). Neither the deployed weights nor the serving stack were independently inspected. MANDATORY DISCLOSURE: measured through the provider's hosted API on 21 Oct 2026; published at the provider's request. THINKING MODE: the API default (no `thinking` field, i.e. "auto"); the row shows one fixed setting, and the provider's docs say that in "auto" mode thinking is skipped when the server is busy, so the default row reflects the server load at the time of measurement (21 Oct 2026, 08:05-08:45 UTC). TRAINING AND SELECTION, provider statement: the public JevBench v1.4/v1.5 items were used as a validation set (recipe and checkpoint choice, serving parameters), not as training data, and the model card says no benchmark test items were used for training; the public set is therefore treated as seen during selection and the normal public-versus-sealed gap rule applies to this row. Provider statements are not independently verified.