What is Jev? TypeSafe AI's decision model for software
What's the deal? TypeSafe AIDealroom has a profile for this one. Try Dealroom → is a San Francisco-based AI lab founded by Diogo AlmeidaDealroom has a profile for this one. Try Dealroom →, a former OpenAI researcher who helped develop RLHF, the technique behind ChatGPT. In September 2026 the lab released Jev, its first System One model, named after the Jevons paradox. Jev is built for software rather than chat.
Developers send it structured questions and it returns typed answers, such as choices or scores, each carrying a calibrated probability that code can act on. Jev does not generate text.
Why now? LLMs have been superhuman at chat for years, but automation has lagged. Sequential generation, mode dropping and overconfidence make chat models slow and hard to trust inside production code. Jev pairs a new architecture with a parallel sampler that answers many questions in one call.
TypeSafe trains it using Reinforcement Learning for Calibrated Decisions. Jev costs $42 per billion input tokens, output tokens are free, and typical responses arrive in under half a second. TypeSafe says Jev is up to hundreds of times faster and cheaper than frontier LLMs, though those comparisons rest on its own evals. Training details stay private. TypeSafe says it used open-weight models and computer-generated synthetic data, according to the Financial Times.
What could go wrong? Jev cannot write prose, so it only fits structured decisions. TypeSafe says answers can still be wrong: its "zero hallucinations" claim means guaranteed schema matching, not perfect accuracy. Jev is in early access behind a waitlist, and the lab has not published weights or parameter counts. General-purpose LLMs already offer structured-output modes. They are a direct alternative. Prices could change.
The signal: Jev spread quickly. Within days it appeared on OpenRouter and on Vercel AI Gateway. Developers built small tools around it for email triage, browser agents and phone agents. Vercel's chief executive reported Jev up to 18 times faster at p95 than a leading LLM.
Deel, a global HR platform, reported up to 59 times lower costs and up to four times faster responses, with quality maintained. TypeSafe closed US$25.9M seed round at US$129.4M valuation more than a year before Jev launched. Almeida said potential investors were "battering down our door". The Financial Times described Jev as a cheap new model taking aim at OpenAI and Anthropic. If decision models such as Jev deliver, more AI value moves from human chat to the software inside everyday products.
Image credit: TypeSafe AI
