Biplov GautamThe category TypeSafe AI named when it released Jev in September keeps growing. This week twenty more decision models and APIs joined the registry, from twenty makers: five hosted APIs, OpenAI's among them, and fifteen open models you can run yourself.
Each page is written from its maker's own card, docs or site, marked as an unofficial listing until the maker claims it, and quotes only the maker's own numbers, with the suite they were measured on. Numbers from different suites don't compare with each other; the comparison table shows every model side by side, suite included.
Hosted APIs
- OpenAI Decisions API. Announced at DevDay 2026 and in limited preview since 29 September, it focuses GPT-6 Luna on questions that have a finite set of answers, with text or images as context. OpenAI has published no pricing or benchmark numbers yet.
- Liquid AI d1. Liquid's first decision model, served only through the Liquid API. It
returns choice, score and yes/no answers with probabilities, on the same request shape as Jev's
/v1/systemone. Its model id isd1:free; Liquid publishes no paid pricing, size or numbers. - Upstage Solar Decide. A System One endpoint on Solar Mini 4, in beta on the Upstage Console at $0.10 per million input tokens, on the same schema as Jev.
- meraGPT Decider 1. $0.03 per million input tokens; meraGPT reports 76.8% on a 2,000-decision typed-decisions benchmark. It is unrelated to Mapika's open Decider.
- Respan Span-01. Behaviour scoring for evals, guardrails and monitoring, at $0.02 per million input tokens, on Respan's own scoring endpoint.
Open models
| Model | Maker | Size | Licence | Maker's own number |
|---|---|---|---|---|
| Xor 1.2 | Juspay | 35B-A3B | Apache-2.0 | 90.0%, JevBench public set |
| JEV-27B | AutoTrust AI Lab | 27B | Apache-2.0 | 88.7%, JevBench public set (family average) |
| StartLux-Decision-4B | StartLux | 4.7B | CC BY-NC 4.0 | 88.3%, JevBench public set |
| this-that-model 1.2 | FLock.io | 1.9B | MIT | 87.8%, FLock's this-that benchmark |
| Jebadiah 27B | Frontier Infra | 27B | Apache-2.0 | 86.6%, JevBench public set |
| JevEmbed | HIT-TMG (Lychee Team) | 4B | Apache-2.0 | 85.9%, JevEmbed-Data test split |
| Bosun 3.1 | Hanno Labs | 2B | Apache-2.0 | 84.9%, DecisionBench (families it trained on) |
| Intern-Decision-4B | InternLM | 4.5B | Apache-2.0 | 80.6%, LocalLLaMA/typed-decisions test split |
| Metask-Jev-4B | Metask Lab | 4.5B | Apache-2.0 | 80.1%, JevBench v1.2 public set |
| NeoHorse-Jev-4B | TokenRhythm | 4B | Apache-2.0 | 75.3%, JevBench public set |
| Standard One 8B | Standard Thinking | 8B | Apache-2.0 | 71.1%, a 2,000-decision typed-decisions suite |
| Lumma-Fev-0.6B |
Five of them, Xor, Standard One, Lumma-Fev, Jebadiah and StartLux-Decision, serve Jev's
POST /v1/systemone, so a client written for Jev switches to them by changing its base URL. They are on the
list of open-source Jev alternatives now.
If you made one of these
Your page is ready for you to claim: write to [email protected] and we'll send a private link that
makes your team its owner. Then you can correct anything on it, publish releases with systemone push,
and keep private repositories for your team. If you'd rather not be listed, say so and the page comes down.
And if you'd like your model to run anywhere with proof that it answers as your own code does, package it with NoulXP, the open standard for System One models.