TypeSafe announced Jev as its first public System One model on 15 September 2026. The launch post frames the product around decisions inside software: supply state, ask typed questions, and use the returned values in application code. This is a different interface from a chat model that generates a response one token at a time.
What developers can try
The documented API accepts a state and a set of questions. Choice selects from defined options, Score evaluates an ordered rubric, and Noul evaluates a yes/no statement. The official introduction says the questions in one call are evaluated independently against the same state. That makes the interface useful for narrow judgments such as routing a ticket or deciding whether a supplied passage supports a claim. The final workflow remains ordinary code.
Read performance claims in context
TypeSafe published speed and cost comparisons in the launch post. Those figures come from its own workflow evaluations and chosen comparison setup; the post also explains why some of the headline improvements may sit toward the high end of real-world gains. They are vendor-reported results, not tests performed by devwithjev. We will not transfer those numbers to individual build listings. If latency matters to your product, measure your own request path, including network time and surrounding application work.
The practical limit
Jev does not generate prose, and its typed result does not decide policy by itself. Developers still need to define the question, validate inputs, set action thresholds, and handle errors. The version-specific jaggedness notes also describe tasks that jev-1.13 handles less reliably, including counting and date comparison. Start with one question in the API quick start, then evaluate it against examples from your own application.
Published 24 September 2026. Product facts checked against the official launch post and documentation available on that date; consult the current docs before implementation.