Decision models (System One)
Decision models read a state (text, JSON, images) plus typed questions and return a probability for every allowed answer in one forward pass — no generated text, nothing to parse. They appeared in volume in Sept–Oct 2026 and already show up in OpenRouter’s usage rankings.
7models tracked
5with open weights
$0.02–0.042hosted input $/1M, output free
2026-10-20checked
Models compared
| Model | Access | License | In / out $ per 1M | Context · input | API | Reported score |
|---|---|---|---|---|---|---|
| Jev 1.14TypeSafe AI 📈 OpenRouter trending | Hosted API | Closed | $0.042 / free | 64K · text | POST /v1/systemone | 61.53 Decision Index 0.3.2 Full (HF Space) |
| Decider V1.2 27B / Decider Flash V1.2 4BPerplexity · 27B (Qwen3.8-27B) / 4B (Qwen3.5-4B) 📈 OpenRouter trending Profile on this site → | Hosted API + open weights | Apache-2.0 | $0.02 / free | 256K · text, image | POST /v1/decisions · aliases pplx-decider-latest / pplx-decider-flash-latest | 65.55 Decision Index 0.3.2 Full (HF Space) |
| Mercury DecideInception 📈 OpenRouter trending | Hosted API | Closed | $0.02 / free | 65K · text, image | POST /v1/decisions (also /v1/systemone schema) | — |
| Clef 27B / Clef-Flash 9BCloudflare · 27B / 9B (Qwen3.8) Profile on this site → | Self-host | Apache-2.0 | self-host | 16K default (encode_record) · text, JSON, image, video | /v1/systemone (Jev-compatible) via SGLang | 53.08 Decision Index 0.3.2 Full (HF Space) |
| d1-3BLiquid AI · 3.1B (LFM2.5-VL-3B) Profile on this site → | Self-host / edge | LFM Open License v1.0 | self-host | 32K · text, JSON, image | transformers · model.system_one(state, questions) | 48.57 Decision Index 0.2.1 (own run of official scorer) |
| JEV-35BAutoTrust AI · 35B-A3B (Qwen3.5-35B-A3B) Profile on this site → | Self-host | Apache-2.0 | self-host | — · text, image | vLLM · POST /v1/decide (≤256 options) | 56.95 Decision Index 0.3.2 Full (HF Space) |
| LayaConvai Innovations · ~0.4B (ModernBERT / mmBERT) 📈 OpenRouter trending Profile on this site → | Self-host | Apache-2.0 | self-host | 512–8K (by checkpoint) · text, JSON | pip laya · Router.predict / POST /v1/systemone (laya-serve) | 4.43 Decision Index 0.3.2 Full (HF Space) |
Scores are self-reported or self-run by each maker on different Decision Index versions (0.2.1 vs 0.3) and are not directly comparable. Prices: TypeSafe docs (Jev); OpenRouter list price (Decider, Mercury Decide). Self-host models have no per-token price.
Decision model vs chat LLM
| Chat LLM | Decision model | |
|---|---|---|
| Output | Free text, code, tool calls | Typed answers with probabilities (yes/no, choice, score) |
| Parsing | You parse the reply; format errors happen | No parsing; the schema is the answer |
| Cost | Input + (more expensive) output tokens | Input tokens only; output free on the hosted APIs here |
| Latency | Seconds (grows with output length) | Milliseconds to ~0.5 s per call, many questions per pass |
| Use for | Writing, coding, reasoning, chat | Routing, triage, moderation, classification, rubric grading, agent guardrails |
| Not for | High-volume labelling at low cost | Anything that needs generated text or explanations |
The shared question types
noulYes / no → probability of yeschoicePick one of named options → probability per optionscoreOrdered rubric level → probability per level + expected score
TypeSafe’s System One schema (state + named questions with type / instructions / criteria) is reused by Clef, Mercury Decide and d1, so switching vendors mostly means changing the endpoint.
- TypeSafe System One
POST /v1/systemone - Perplexity Decisions
POST /v1/decisions - Inception Decisions
POST /v1/decisions - Laya (self-host, Jev-compatible)
POST /v1/systemone
Notes per model
- Jev 1.14The first “System One” model; defines the API others copy. Output tokens free. Rate limits currently adjusted dynamically. Docs (2026-10-19): alias jev-latest → jev-1.13.0 (stable), jev-preview → jev-1.14.0; both $0.042 per 1M input tokens. [1][2][3]
- Decider V1.2 27B / Decider Flash V1.2 4BV1.2 (HF repos created 2026-10-15; on OpenRouter 2026-10-16) reads all questions about one state in a single sequence, so grouped questions cost about as much as one. Official price: $0.02 (V1.2) / $0.01 (Flash V1.2) per 1M input tokens, output free. Open weights, Apache-2.0. Model card: Decision Index 0.3 public 68.09 (joint requests) vs 62.30 for V1.1. [1][2][3][4][5]
- Mercury DecideUp to 14 decisions/s per Inception; free rate-limited endpoint on OpenRouter. Released 2026-09-30. $0.02/1M input is a 50% launch discount (regular $0.04; output not billed; no end date published, checked 2026-10-20). Vals AI’s independent test: on par with Jev on claim verification, ahead on LegalBench. [1][2][3]
- Clef 27B / Clef-Flash 9BJev/SystemOne-compatible; median latency 209 ms (Clef) vs 38.8 ms (Clef-Flash) in Cloudflare’s internal run. [1][2][3]
- d1-3BBuilt for edge: ~8 ms per decision on an RTX 4090, ~30 ms on Apple M5 Pro, ~50 ms on Jetson Orin Nano (Liquid’s numbers). [1][2]
- JEV-35BIndependent open model, not affiliated with TypeSafe. NVFP4 variant (25.6 GB) runs on one 32–48 GB GPU. Own public-suite score 59.73; on the 0.3.2 Full board 56.95 (#37). [1][2]
- LayaNon-autoregressive System 1: ~33 ms/pass, typed answers with calibrated probabilities; English / multilingual / typed-decisions checkpoints. Claims Jev-compatible HTTP shape. On Decision Index 0.3.2 Full at 4.43 (#164 on the 2026-10-16 board). [1][2][3]
Leaderboards & independent tests
- Decision Index 0.3.2 (HF Space)Verified from data/index.json (generated_utc 2026-10-16T16:39:08Z, label Decision Index 0.3.2; 173 models, +8 vs 2026-10-15). New #1 StartLux-Decision v1.1 27B 69.12; Perplexity Decider v1.2 65.55 (#4); new Amos 1 Large 64.08 (#6), Lychee-27B 62.99 (#8), KnowLine 35B-A3B 62.92 (#9), span-02 62.56 (#11). Jev 1.14 reference 61.53 now #13. Open-weights with weights_repo in top 50: Surogate Rune 26B-A4B v3 (#35), Eikos-27B (#46).
- #1 StartLux-Decision v1.1 27B · 69.12
- #2 vLLM-SR d3 (27B) · 66.26
- #3 linne-oak 1.0 (31B) · 65.61
- #4 Perplexity Decider v1.2 (27B) · 65.55
- #5 Decider 31B (NVFP4) · 64.62
- #6 Amos 1 Large (27B) · 64.08
- #13 Jev 1.14 (hosted reference) · 61.53
- Vals AI — Jev / Mercury Decide evaluationsIndependent claim-verification and LegalBench tests
Sources: vendor docs, Hugging Face model cards and OpenRouter public pages, checked 2026-10-20; OpenRouter usage as of 2026-10-19 (CC BY 4.0). First version of this topic — more models and our own tests to follow.