Jev: TypeSafe's System One decision model
Jev is not a coding agent, but TypeSafe AI's System One model. Typed probabilities instead of chat — and what is public about cost and usefulness.
Status: 18 September 2026. Jev is TypeSafe AI’s first System One model, announced on 15 September 2026. It is in early access. It is not a chatbot and not a coding agent.
What Jev is
TypeSafe describes Jev as a function call for software: unstructured state in, typed probabilistic decisions out. You send program state plus questions; you get structured answers the surrounding code can branch on. There is no generated prose to parse.
The official docs define three question types:
- Choice — pick one option from a set you defined, with probabilities and confidence
- Score — place the state on an ordered rubric
- Noul — return a yes/no probability between 0 and 1
Questions in one request are evaluated in parallel against the same state. The TypeScript and JavaScript client is @typesafe-ai/sdk.
Cost: cheap enough for the request path
The published list price is $0.042 per million input tokens, with output free. TypeSafe states that figure on the launch post and on the model card; Vercel’s AI Gateway lists the same input rate. TypeSafe also quotes end-to-end response times of 70–500 ms.
Those numbers are the vendor’s. Independent production measurements are still thin. Even taken at face value, they put Jev in a different cost class from chat models, whose published input prices TypeSafe places between $0.20 and $10 per million tokens.
How useful is it now?
Useful where software already owns the workflow and needs a judgment: classify a ticket, route a request, score risk, gate an agent step, or check a guardrail. Vercel and Netlify now expose Jev through their AI Gateways for that kind of call.
It is not useful as a replacement for a coding agent or a chat interface. Jev does not write replies, code, or explanations. Early access still applies on TypeSafe’s own API. Probabilities still need thresholds you set and check against your own labeled examples.
Anyone looking for a cheap decision primitive can try it. Anyone expecting a general assistant will be looking at the wrong tool.