He Says He Co-Invented ChatGPT. His New AI, Jev, Won't Write a Word.

Half true, and the half that matters. Because outputs are enumerated in advance, an off-schema answer cannot come back, so the parser-and-retry layer genuinely disappears. But strict structured-output modes on the major LLM APIs already return a valid enum with no parser. What is added, on TypeSafe's account, is a calibrated probability, and no algorithm or architecture has been published for it.

Direct TypeSafe keys were still early access when this was written; the practical route was AI Gateway with AI SDK 7 as the client: pnpm add ai@latest @ai-sdk/typesafe-ai, then experimental_evaluate with an evaluation model. If you need the direct endpoint, confirm current access before designing around it.

The provider documents retries for 429 and 529 up to maxRetries, which defaults to two. Two retries is thin for anything on a user-facing path, so set it deliberately and wrap the call in your own timeout and circuit breaker. Nothing in the guide documents rate limits or sustained throughput per key.

No. Confidence arrives for Choice and Score only; Noul returns just the 0-to-1 probability of yes, with no separate confidence field. Values are also rounded to two decimals, which matters if you plan to threshold tightly or average many calls. And the quoted 0.114s at $0.000081 against 8.566s at $0.013880 is a vendor example, not a measurement you can assume.

Run pnpm add ai@latest @ai-sdk/typesafe-ai, then export TYPESAFE_AI_API_KEY and call experimental_evaluate with an evaluation model; the AI SDK's TypeSafe provider is the documented path.

Choice picks one option from 1 to 255, Score places the input on 2 to 10 ordered levels and returns a fractional score, and Noul is yes/no answered with the probability of yes, which the AI SDK calls boolean.

Through Vercel AI Gateway, which its changelog says Jev joined on 16 September with AI SDK 7 as the client; direct TypeSafe keys were still early access, so the gateway was the practical route.