Vercel puts Jev behind a plain HTTP endpoint that returns probabilities instead of text

Vercel Changelog·medium signalXBlueskyLinkedInCopy link

AI Gateway now exposes TypeSafe AI's Jev decision model three ways: point an existing TypeSafe client at the /typesafe endpoint, POST to /v1/evaluate, or call experimental_evaluate in the AI SDK. You send state plus a questions object and get typed answers back with probabilities attached — boolean returns 0 to 1, choice selects from named options, score rates against a defined scale — which removes the parse-and-validate step that wraps most LLM classification calls today. All three paths bill through AI Gateway, so decision-model spend shows up in the same usage and observability views as token spend.