Jev is TypeSafe AI's first "System One model": a model that doesn't generate text. It takes a state and a list of typed questions and returns typed values—a yes or no, an option from a list, a score—with calibrated probabilities. It launched in early access on September 15, 2026, and within a week it became the main topic among those of us building systems with LLMs. I went through the documentation and the first independent analyses to answer what matters in practice: what it is, how to call it, what it costs and where in a real system it makes sense to use it.
TL;DR
- Jev doesn't write: it answers
choice,scoreornoul(yes/no) questions with a value, its probabilities and a confidence level your code can use directly.- It costs 0.042 USD per million input tokens, output is free and the published latency is 70 to 500 ms. It's for classifying, routing, prioritizing and evaluating, not for writing or reasoning.
- Use it as a fast decision layer in front of an LLM: if confidence is high you act, if it's low you escalate to the large model. No one has independently reproduced the benchmarks yet.
What is Jev and how is it different from an LLM
An LLM generates text token by token. When you use it to decide something—is this ticket urgent? which department does it go to?—you ask for a JSON answer, parse it and validate it in case the model changed the format. A good part of the code around an LLM in production exists only to turn prose into a value the program can use.
Jev removes that step. You send it a state (the context: a message, a document, an agent's history) and a map of questions, each with a declared type. What comes back isn't text, but a value of the type you asked for plus a probability distribution. There's nothing to parse, and the model can't invent a category that doesn't exist, because it can only choose among the ones you defined.
TypeSafe AI calls it "System One" after the distinction between fast, intuitive thinking (system 1) and slow, deliberate thinking (system 2). The idea is that current LLMs cover system 2, and that a large share of the decisions a piece of software makes don't need deliberation, just a fast and measurable judgment. The company is based in San Francisco, was founded by Diogo Almeida—a former OpenAI researcher and one of the co-authors of RLHF—and announced a 40 million USD seed round led by DCVC alongside the launch. The demo that spread the most was Jev playing Doom: it receives a frame and returns the next move, with no commentary.
The three question types: choice, score and noul
The whole API comes down to three primitives. Every question you ask Jev is one of these types:
| Type | What it answers | What it returns | Example |
|---|---|---|---|
choice |
One option from a closed list | The chosen option, its confidence and the probability of each option | Which department does this ticket go to? |
score |
A level on an ordered scale of 2 to 10 levels | The score, its confidence and the distribution per level | How frustrated is the customer, from 1 to 5? |
noul |
Yes or no | The probability that the statement is true, between 0 and 1 | Is the message asking for a refund? |
The difference between the answer and the confidence is the most useful part of the design. The answer says what the model chose; the confidence, computed from the shape of the distribution, says whether you should act on it. If the probability is concentrated on one option, confidence is high. If it's spread across two or three, it's low, even when the winning option is the same. An LLM has no native equivalent of that signal: it returns a category with the same apparent certainty whether it's right or unsure.
Keep reading
That is the first half. The full walkthrough — with the rest of the implementation, the trade-offs and the things that only show up in production — is on my blog:
Read the full post on ramonchancay.me →
Originally published at www.ramonchancay.me/blog/what-is-jev-typesafe-ai.
Top comments (0)