Jev no longer requires AI SDK: call evaluation with one POST from any language

Jev is an evaluator from TypeSafe. You give it a conversation state and a list of questions; it returns an answer: yes or no, a choice from options, or a score on a scale.

There are now three entry points. The HTTP API at /v1/evaluate opened on 16 September, and the TypeSafe-compatible one on the 22nd. Evaluation can be embedded into a backend in any language, not only a JS project.

Migration takes two edits. In an existing TypeSafe client, change baseURL to https://ai-gateway.vercel.sh/typesafe and the key to AI_GATEWAY_API_KEY. POST /typesafe/v1/systemone and GET /typesafe/v1/models are supported.

There is one gotcha. In the TypeSafe-compatible API, a boolean question is called type "noul", and the response arrives in the noul field. In the HTTP API and AI SDK, the same type is called "boolean", and the field is probability.

Raw HTTP consists of three parts: a POST to /v1/evaluate, a header with the AI_GATEWAY_API_KEY key, and a body containing model "typesafe-ai/jev", state, and questions. Jev does not respond through OpenAI-, Anthropic-, or Cohere-compatible endpoints.

The price is $0,042 per million input tokens, with output not billed at all. Evaluations are free under a promo until 25 September 2026, but a card must still be attached to the team.

The request cap is 64 000 context tokens, including 32 000 for state. The rate limit is strict: a third-party jevaiguide test got 131 questions through in roughly an hour and a half (19.09).

Jev launched on 15 September, and AI Gateway adopted it faster than any previous launch: 10% of teams in 18 hours, 13% of paid teams in a day. That is twice the rate of the GPT-5.6 family and 6 times that of Fable 5.1. TypeSafe promises up to 193,6 times faster workflow evaluations at up to 444,6 times lower cost than LLMs.

The promo ends on 25 September 2026: after that, each million input tokens will cost $0,042.