What is Jev?

TypeSafe AI came out of stealth this week with a model called Jev, and the line that stopped me was that it does not generate text. Not "it can output JSON if you ask nicely" — it cannot write a sentence at all. That sounded either useless or quietly clever, so I asked claude to build me something to press until I knew which.

Open the explainer

Here is where I landed. When I use claude or a local model, I am the one reading the answer, so prose is the right shape. But most software using AI is not reading anything — it wants one fact so it can pick a branch. Getting that out of an LLM means asking for a decision, receiving a paragraph, and then parsing the paragraph back into the decision. The model is right and the packaging breaks.

Jev skips the packaging. You send it some state and a named set of typed questions, and you get back one answer per question:

• Noul — a yes/no, returned as a probability between 0 and 1.
• Choice — one option out of a list you defined, plus the probability of every other option.
• Score — where the state sits on a rubric you wrote in words, as a number between your levels.

Three things about it I keep thinking about:

• The answer cannot be a type error. A Choice can only ever come back as one of your own keys, so there is no schema to validate and no retry loop when the JSON comes back malformed.
• You always get the probabilities, and on Choice and Score a confidence number derived from them. So the model can say "I don't know", and your code can route on that — act, confirm, or send it to a human — with a different bar for a refund than for showing a page.
• Questions are evaluated in parallel against a state that is read once, so asking thirteen instead of one barely changes the response time. You end up asking everything speculatively and throwing away what turned out to be irrelevant, which is the opposite of how you'd budget an LLM call.

It is quick and cheap enough to feel like a different category: 70–500ms, $0.042 per million input tokens, and output tokens free. Cheap enough to ask on every row, every retrieved passage, every message.

The honest half: it cannot count, it cannot do arithmetic, it reads dates as text rather than as ordered things, it is text only, and English is where it is strongest. TypeSafe publish all of that themselves on a page they call jaggedness, which I liked them more for. It does not replace an LLM any more than WebMCP replaced MCP — it sits in front of one, or behind one, doing the fast unglamorous judgements.

Early access opened on 15 September 2026, so this is me reading the docs rather than me having shipped anything with it.