Jev is a decision model, not a chat model
Jev is TypeSafe AI’s first public System One model, announced on 15 September 2026. You send the state your software already has — a ticket, a log line, a tool call, a retrieved passage — plus questions whose answers are bounded in advance. What comes back is a typed value, not a paragraph.
That is the product. TypeSafe trained Jev to make fast, structured judgments that code can consume directly. It is a poor fit for drafting, summarizing, or explaining, and TypeSafe says the current version is not trained to generate text.
What Jev returns: choice, score, and noul
Every question is one of three types. A choice picks one label from a set you define, up to 255 options, and returns the winner plus a probability for each option. A score places the state on an ordered 2–10 level rubric and can land between levels. A noul answers a yes/no question with a single probability from 0 to 1.
Several questions can share one request and are evaluated against the same state. Your application still owns the side effect: Jev can select “billing”, but it does not issue the refund.
- Current public model string: jev-1.13.0, also exposed as jev-latest and jev-preview.
- Vendor latency claims for System One tasks are typically 70–500ms; measure your own path.
- Input is billed at $0.042 per million tokens. Output is free because the response is a typed value, not generated text.
Where the Jev model is a good fit
Start with narrow, repeated judgments: support triage, model routing, moderation gates, lead scoring, citation checks, and “should this tool call run?”. Keep open-ended writing, deep research, and long reasoning with a general-purpose language model.
The usual production pattern is hybrid. Jev decides the branch; GPT, Claude, or another LLM writes whatever a person will read. That is also why “Jev vs GPT” is the wrong framing if you expect one model to replace the other.
Who makes Jev, and what System One means
TypeSafe AI is a San Francisco lab that came out of stealth with a $40M seed round led by DCVC. The company names the category System One after fast, intuitive System 1 judgments — the cheap branches software makes all day — rather than slow System 2 reasoning.
TypeSafe also describes a training method it calls Reinforcement Learning for Calibrated Decisions (RLCD). Treat vendor speed, cost, and “zero hallucination” claims as claims about schema-bounded output, not as proof that the selected label is always the correct one.
FAQ
What is Jev?
Jev is TypeSafe AI’s first System One model. You send application state plus typed questions; it returns a choice, a score, or a yes/no probability your software can branch on. It does not write prose, code, or explanations.
What is Jev AI / the Jev model?
Jev AI is the public name for TypeSafe’s hosted decision model, currently published as jev-1.13.0 with aliases jev-latest and jev-preview. It is built for routing, triage, scoring, and gating inside software, not for chat.
Is Jev open source?
No. Jev is a closed, hosted API. TypeSafe has not released the weights, architecture, or training recipe. Community projects such as openjev and jevlike experiment with similar interfaces on open models, but they are not Jev.
Is Jev free to use?
There is no published free tier from TypeSafe. Input tokens are billed at $0.042 per million and output is free. Some independent playgrounds and the Vercel AI Gateway let you try the model under their own limits.
How do I get access to Jev?
Create an API key in the TypeSafe console at console.typesafe.ai, or call the model through Vercel AI Gateway as typesafe-ai/jev. The public endpoint is POST https://api.typesafe.ai/v1/systemone.
Can Jev generate text?
No. TypeSafe documents that jev-1.13 is not trained to generate text. Use Jev for the branch, then a general LLM if a person needs a written reply.