Jev by TypeSafe: a structured decision model, not a chatbot
Jev is TypeSafe AI’s first System One model. It returns typed decisions and probabilities instead of generated text. This page is a model profile — what it is, what it is for, and what the claims still need.
What is Jev
Jev is TypeSafe AI’s first public System One model. You give it a state (text, JSON object, or array of text) and a set of typed questions. It returns bounded answers plus probabilities — a yes/no mass (noul), a distribution over options you defined (choice), or a position on a rubric (score).
It does not chat. It does not write. TypeSafe’s launch line is that it “gives up string generation” so software can consume the answer directly. That is a product claim, not a lab result from this site.
Developer / company
TypeSafe AI. The public founder story ties the company to former OpenAI research (including RLHF-era work) and a training method they call reinforcement learning from calibrated decisions (RLCD). Treat biography as background, not a quality score.
Launch date and current version
Public launch landed in mid-September 2026. Third-party catalogs list jev-1.13 / jev-1.13.0 around 18 September 2026. Pin versions in production; early System One APIs are still moving.
Pricing (vendor claim)
Published rate: $0.042 per million input tokens, output free (“too cheap to meter”). TypeSafe itself notes it cannot prove the price is unsubsidized. See the pricing review for the comparison intent. Official rate cards belong on a Jev specialist site, not here.
Input / output
- Input: text-shaped state. Vendor docs mention strings, JSON objects, and arrays of text. Images, audio, and video are called out as not supported yet.
- Output: typed decisions and calibrated probabilities — not assistant prose.
- Context numbers circulating in secondary writeups (for example 64k per request) are unverified by this desk.
Main use cases
The honest list is short: classify, route, score, gate, extract a closed field. Agents and RAG only make sense if Jev sits beside a generator. Chat, marketing copy, and code generation are the wrong category.
API availability and integrations
Vendor: Decisions / System One API. Third-party: listed on OpenRouter as Jev 1.13 and discussed on Requesty as typesafe/jev-1.13.0. That is catalog availability, not an integration maturity score.
Strengths
- Job is narrow and therefore easy to evaluate later.
- Published input price is far below typical frontier LLM input+output bills if you only needed a decision.
- Bounded outputs remove a class of “model wrote a poem instead of a label” failures — according to the vendor’s type story.
Limitations
- No generation. If the user needs an explanation, you still call an LLM.
- Most speed and “cannot hallucinate” language is still a vendor claim.
- Calibration is a group statistic. It does not make any single answer true.
- Ecosystem is early. Few battle-tested production stories are public.
Latest updates
September 2026: public launch, jev-1.13 catalog listings, first independent-adjacent monitor writeup. This profile will add a dated changelog when versions or prices move — not a tutorial feed.
Task fit, not a winner
| Task | Jev | Typical LLM |
|---|---|---|
| Classification | Structured decision with a bounded label set | Possible, usually via generated text or JSON schema |
| Model routing | Native fit: pick a route, return a typed choice | Possible, but you pay generation cost for a decision |
| AI agents | Use as a judge / router / gate, not as the actor | Use as planner, writer, tool-caller, and explainer |
| RAG | Score, filter, or route retrieved chunks | Synthesize an answer from retrieved context |
| Chat | Not the job. No free-form conversation | Primary job |
| Writing | Not the job. No prose generation | Primary job |
| Coding | Not the job, unless you only need a pass/fail or route | Generate, explain, and iterate on code |