TypeSafe AI's Jev lands on AI Gateway: decisions without token-by-token text, up to 193.6x faster in vendor tests
Synopsis
TypeSafe AI's probabilistic decision model Jev is now available on Vercel AI Gateway, letting an application hand in a state and get back typed Choice, Score and Boolean answers with probabilities. Regular language models generate text one token at a time for the application to parse and validate; Jev evaluates all declared questions in parallel and returns structured answers directly. TypeSafe reports Jev was up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations, for uses such as picking the next tool in an agent loop, deciding whether to continue or stop, scoring risk, and verifying outputs.
Interpretation
Applications can now call TypeSafe AI's Jev through AI Gateway, turning "what should happen next" into a structured query instead of asking a language model to write text that is then parsed. Regular language models generate text one token at a time, and the application then parses and validates it; Jev evaluates all declared questions in parallel and returns typed answers plus probabilities directly. TypeSafe reports Jev was up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations.
A caller supplies a shared state and a set of named questions and receives several decisions in one call. AI SDK 7 exposes Jev through the experimental evaluate API, where a call specifies the model, the state and a map of questions, with the model identified as typesafe-ai/jev. The example turns one support case into queue, priority and refund-review decisions, and routes uncertain cases to manual review.
Perspective
The demonstrated use is calling Jev as typesafe-ai/jev through AI Gateway, submitting a state and named questions through the experimental evaluate API in AI SDK 7.0.105 and later, and receiving answers; the example scenario is routing, priority and refund-review decisions for one support case.
The speed and cost multiples come from TypeSafe's self-reported workflow evaluations, and the material does not describe the evaluation set or task details; it also advises calibrating probabilities and confidence against labeled examples from the workflow, so returned probabilities need calibration for a given setting before use.