The short version
Jev is the first public model from TypeSafe AI, announced on 15 September 2026 as the first member of a new category the company calls System One models.
Instead of writing sentences one token at a time, Jev reads a block of unstructured state and answers a set of typed questions against it. You get back values and probability distributions — not prose you have to parse.
TypeSafe describes it as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.
Why this matters
Almost every AI product you use today is built on string generation. That is a strength for chat and a liability for software: to act on a decision you must coax a text model into a schema, parse it, validate it, and still accept that it might return something the parser rejects.
Jev removes that layer. The set of possible answers is declared before the call, so a type error is not unlikely — it is impossible by construction. What you lose is prose. What you gain is a model you can put inside a latency budget.
The name, and the idea behind System One
TypeSafe named the class after Daniel Kahneman's Thinking, Fast and Slow. Kahneman's System 1 is fast and intuitive; System 2 is slow and deliberate. Chain-of-thought models live firmly in System 2.
“Jev” comes from William Stanley Jevons, the economist who noticed that cheaper steam engines increased total coal demand rather than reducing it. TypeSafe's bet is that every order-of-magnitude drop in the cost of intelligence exposes far more use cases than it replaces.
For the wider category, read what a System One model is.
What Jev actually returns
TypeSafe exposes three question types, and you can mix them in one request: Choice (pick from a list), Score (rate on a rubric) and Noul (roughly, how true is this statement, from 0 to 1).
Every question is answered in parallel against the same state, and every answer carries a confidence or probability. Adding questions barely changes response time, because the model is not generating a longer string — it is resolving independent decisions. See Choice, Score and Noul.
- Choice — a category plus a probability for each option
- Score — a numeric score on a rubric, with a legend
- Noul — a 0-1 truth estimate for a single claim
- Confidence — how sure the model is, per question
Where it fits in a real stack
Jev is the wrong tool for chat, code generation, long-form writing, or anything a human needs explained in words. It is the right tool for the decision that the surrounding code makes hundreds or millions of times: classify, route, score, extract a judgement, branch where hand-written rules are too brittle.
The economics are the interesting part. TypeSafe lists $0.042 per million input tokens with free output and 70-500ms end-to-end latency. That is the difference between a model you call in a background job and a model you call inside a request.
Common mistakes
These show up in almost every first integration with a decision model.
- Treating Jev as a cheaper chatbot — it cannot write your caption or your email
- Asking one giant question instead of decomposing it into atomic ones
- Ignoring the confidence number and acting on every answer equally
- Expecting a rationale: Jev returns a value, not a reason
How PixaSocial Ai helps
PixaSocial Ai is built as one workspace for personas, planning, creatives and publishing, which is exactly the shape of problem decision models are good at: many small judgements over shared context.
Start with the articles hub for the rest of this cluster, or register to see how audience models and content workflows are wired together.