Updated on September 18, 2026
Jev is the first 'System One' model from TypeSafe AI, a decision engine that returns typed answers and calibrated probabilities instead of text. Like an LLM it reads plain language, but it never writes any back. Developers wire it into software to sort, route and score at high volume. TypeSafe opened early access in September 2026 behind a waitlist, with billing applied to input tokens only, output costs nothing.
Jev takes a state as input, plain text, JSON objects or arrays, evaluates typed questions in parallel and returns every answer within 70 to 500 milliseconds. No prose, no code, no written reasoning.
Paste in a support ticket, ask at once for the department, the urgency level and the odds of a refund request, and all three verdicts land together, each with its own confidence score. Training relies on a method called RLCD, aimed at calibrated probabilities, meaning that across many predictions rated 80 %, roughly 80 % turn out correct (calibration describes averages, not any single answer, so test it against your own data first).
| Primitive | What Jev returns | Sample use |
|---|---|---|
| Choice | One option out of up to 255 | Routing a ticket to the right team |
| Score | A position on a 2 to 10 level scale | Grading the severity of a security alert |
| Noul | A yes/no probability | Flagging a jailbreak attempt |
TypeSafe AI came out of two years in stealth with $40 million in seed funding led by DCVC, and founder Diogo Almeida is a former OpenAI researcher who co-authored the work behind InstructGPT. As for the name, Jev is a nod to the Jevons paradox, the idea that a cheaper resource ends up consumed in far greater quantities.
On TypeSafe's internal test bed, four workflows including incident response and invoice handling, Jev reaches about 68 % agreement with reference answers, on par with mid-tier LLMs at a fraction of the time and price. One demo even has it playing Doom at ten decisions per second. Yes, Doom.
In production it mostly runs as a cascade among AI automation building blocks, where Jev classifies and routes each case, your code handles the obvious ones, and a general LLM only sees the ambiguous leftovers.
Jev is in early access, sign-ups go through a waitlist at typesafe.ai and, according to early reports, keys arrived within a day or two. The console includes a Playground with guided lessons and worked examples, resume screening, support-agent audits, ticket routing.
Pricing sits at $0.042 per million input tokens, with output free of charge. Those are the launch rates, and the page published by TypeSafe is the only one that counts on the day you open an account.
No, Jev bills per use on input tokens only, and output is free. No free-credit program was described at launch, though TypeSafe puts the cost at fractions of a cent per decision. Getting in currently means joining the early-access waitlist on the official site.
Within its schema, yes. The model cannot return a value outside the options you define, so no malformed JSON and no invented label. A typed answer can still be wrong on substance, and a bad choice carrying a valid probability remains an error worth monitoring.
Jev complements LLMs rather than replacing them. It writes no text and no code, it settles closed questions. The recommended setup is a cascade, Jev classifies and routes cheaply, your code handles clear cases, and a general model only receives the hard minority.
Text only, as strings, JSON objects or arrays. Images, audio and video are not supported. TypeSafe also advises sending only the state relevant to the decision, since accuracy drops when the context fills up with unrelated material.
Verdict: For sorting, scoring and routing entire streams of tickets, logs or agent runs, Jev adds a decision layer priced at a sliver of what generative models charge, and teams that keep a general LLM around for the tricky minority will get the most out of it.
