TypeSafe AI launches Jev decision model with a ticket-routing demo

On October 9, 2026, Juejin published a walkthrough of Jev, the first "System One" model from San Francisco startup TypeSafe AI. According to the report, Jev does not generate text; it only makes judgments. The company positions it as "Decisions, not strings," describing it as an intelligent if-statement embedded in code: the model handles fuzzy semantic judgment while the code handles precise rules, permissions, and actions. The name System One comes from Kahneman's Thinking, Fast and Slow and refers to fast, intuitive judgment. Jev is not an acronym but a nod to economist William Stanley Jevons, invoking the Jevons paradox: once the cost of a single intelligent judgment drops very low, places in software where calling an AI once seemed uneconomical will start using one. The report says Jev suits high-frequency, small decisions such as ticket routing and content moderation, answers only three kinds of questions, and can compute them in parallel within a single request, so asking 1 question and asking 4 questions have nearly the same latency.

The report cites official comparison figures: in a specific workflow evaluation, Jev was up to 193.6x faster and 444.6x cheaper than a comparison large model. It notes the test tasks were designed by the company's own team and the reference answers were generated by other large models rather than labeled by humans, and that the company itself acknowledges the numbers are on the high side; the author advises treating them only as directional. The article also states that within 24 hours of launch, nearly 13% of Vercel paid teams had adopted Jev, making it the fastest-adopted new model in Vercel platform history.

The body of the report is a beginner tutorial: starting from applying for an API Key, it walks readers through sending requests and running a complete intelligent ticket-routing demo. It first spends a few minutes explaining what Jev is, how it differs from traditional large models, and what it is good for, then moves to hands-on steps. The verifiable progress so far rests on this single report describing the launch and the tutorial.