A model with no prose in it
TypeSafe AI, a startup founded by the former OpenAI researcher Diogo Almeida, opened early access on 15 September to a model that does not generate text. Jev takes unstructured input and a set of typed questions, and returns values and probabilities — a classification, a route, a score — and nothing else.
The company calls this a System One Model, and says it is the first of its kind: a model “built to make fast, structured decisions that software can use directly” rather than one optimised for a human reader. Almeida described it to AI News as “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out”.
TypeSafe spent two years in stealth on a $40m seed round before the release.
The numbers the company gives
TypeSafe’s own figures put Jev’s end-to-end response time at 70 to 500 milliseconds, against 3 to 329 seconds for frontier language models on the same work, and price input at $0.042 per million tokens with output free — “too cheap to meter” — where conversational models charge roughly $0.20 to $10 per million input tokens. AI News reported evaluations showing Jev up to 193.6 times faster and 444.6 times cheaper than language models on narrow decision tasks.
These are vendor numbers, and TypeSafe says so itself. Its announcement notes that the workflow evaluations “are on the higher end of real world gains” and that “some bias could exist” because the workflows were written by its own capabilities team. No independent benchmark of Jev has been published.
How it is built
Rather than predicting one token after another, Jev produces its typed answer in a single parallel query. The training method is what the company calls Reinforcement Learning for Calibrated Decisions, aimed at making the probability attached to each answer mean something rather than merely rank options.
The structural consequence TypeSafe leans on hardest is that the output is constrained to a schema the caller defines, so the model cannot return a value outside it. The company says this means Jev “can’t hallucinate”. That is true in the narrow sense that it cannot invent a type — it can still return the wrong option with high confidence, which is a different failure and one only independent testing will size.
Why it is not a chatbot competitor
Jev cannot write an email, summarise a document or explain an error. That rules out most of what people buy language models for, and is the point: a large share of production LLM calls are not writing tasks at all but routing, filtering, scoring and judging, where generating prose token by token is overhead. TypeSafe is selling into that slice, and the demonstrations reflect it — a bot playing Doom from live game state, and navigating between Wikipedia articles.
What to watch
Two things. The first is an independent evaluation, without which the speed and cost claims remain the company’s own. The second is whether “System One Model” turns into a category other labs build for, or stays the name one startup gave its product.
