TypeSafe AI launches Jev, a hosted model for typed software decisions

TypeSafe AI introduced Jev, which it calls its first System One model, in a post dated 15 September 2026. The page metadata shows 19 October, so the date is inconsistent. Jev takes unstructured state and typed questions and returns typed outputs with calibrated probabilities and a confidence estimate. It does not generate free text.

The post lists classify, route, score, extract, branch, map-reduce and verify as uses, and says Jev supports up to 255 choices per decision. Access is early access: TypeSafe is onboarding developers from a waitlist, with docs at docs.typesafe.ai and a console at console.typesafe.ai.

The post gives a price of $0.042 per million input tokens with free output. It claims 70 to 500 ms end-to-end latency. It also states that its speed tests ran from a laptop, that reference answers average two external models, and that it cannot prove the pricing is not subsidised. Licence and weights are not stated.

A third-party write-up reports a $40 million seed round. A Classmethod test found lower latency and cost for a narrow routing case, but did not measure accuracy broadly.

This is the model that other decision releases measure themselves against. The post is candid about its own caveats, which is rarer than the speed claims.

Sources

• https://typesafe.ai/blog/introducing-system-one-models-and-jev
• https://typesafe.ai
• https://letsdatascience.com/news/typesafe-ai-launches-jev-decision-model-889a38c0

Editor's first pass (AI-assisted)

• Real 3: Access is early access with a waitlist; weights are not stated, and the docs and console are the only way in.
• Useful 4: Classify, route, score and verify calls with calibrated probabilities are a real software task.
• Notable 4.5: Most other decision releases this autumn cite Jev or copy its API contract.
• Novel 4: It set the format; the launch post calls it a function call with typed probabilistic output.