TypeSafe AI came out of stealth with $40 million in seed funding and a model that works nothing like the chat you’re used to. Jev doesn’t answer with text. It answers with a typed decision.
What Jev returns
Instead of a string you then have to parse and validate, Jev hands back values of a fixed type: a yes/no probability, a pick from a defined list, or a score on a specified scale. Every decision carries probabilities and a confidence measure. That lets an application set thresholds — high enough and the workflow proceeds on its own; ambiguous and the case goes to a person.
The company’s example is insurance. Jev looks at the available evidence about a property and estimates the likelihood it catches fire. High confidence and the review runs through; a borderline case lands with an underwriter.
TypeSafe calls this a «System One Model», after Kahneman’s fast thinking, and the training method «Reinforcement Learning for Calibrated Decisions». The goal explicitly isn’t conversational quality, it’s epistemically honest probabilities. All outputs are produced in one query, in parallel, rather than token by token.
Who’s behind it
Founded in 2024 in San Francisco by Diogo Almeida, who worked on RLHF, InstructGPT, ChatGPT and GPT-4 at OpenAI. Co-founders are Erik Gafni and Sasha Sheng. DCVC led the round; Forbes reports a $200 million valuation, citing a person familiar with the deal. Jev is in early access behind a waitlist.
Almeida on today’s models, speaking to Forbes: «We’ve been optimizing for humans, and we’re superhuman at pleasing humans.»
The numbers nobody has checked yet
Under 100 milliseconds per answer, up to a hundred times faster and cheaper than frontier models — that’s TypeSafe’s claim. The pricing page lists 39 cents per thousand workflows against $3.31 for GPT-5.6 Luna and a comparison figure for Anthropic’s Claude Haiku 4.5. The company’s own tests put Jev at nearly 194 times faster and about 445 times cheaper.
None of that is independently verified, and all of it depends on workload, network location and how the comparison is drawn. The claim that Jev can’t hallucinate holds only for the shape of the output: a typed value is always well-formed, which doesn’t make it correct. Whether the confidence scores track actual accuracy is something you have to measure on your own data.
A control layer beside the language model
Jev is interesting not as a rival to Claude but as a component next to it. TypeSafe names as its biggest opportunity exactly the high-volume processes where agents are expensive and shaky today: classifying service requests, evaluating invoices, triaging security alerts — and reviewing the output of AI agents.
That last one is the real one. Anyone who has built an agent that uses a second language model to decide whether the first one is finished knows the problem: the reviewer costs as much as the worker and is just as badly calibrated. A small, fast model that only says «yes, 0.91» would be the obvious fix. Which leaves the question that sits precisely where TypeSafe wants to win: can you trust the confidence score? No pricing page answers that.
Sources: