TypeSafe leaves stealth and ships Jev, a model that refuses to chat
Former OpenAI researcher Diogo Almeida posted a two-year stealth project on X — a System One model named Jev, built for typed software decisions instead of generated text.
On September 15, 2026, Diogo Almeida — who describes himself as a co-creator of RLHF / ChatGPT work at OpenAI and now CEO of TypeSafe AI — posted a launch thread that quickly became the public origin story for Jev.
The question he put at the top of the thread was not “how do we make a better chatbot.” It was why superhuman chat models had not produced the automation wave people keep promising. His answer, after two years in stealth, is a different product shape: a model that software can call like a function.
Jev is TypeSafe’s first public System One model. In the company’s telling, you send unstructured state (text, program data) plus a set of questions whose answers are already typed — a choice among options, a score, a yes/no. The model returns those values with calibrated probabilities. It does not write a paragraph, a tool call in prose, or a JSON blob you have to parse and hope is valid.
The launch thread’s headline claims, which TypeSafe repeats in its blog post, are:
- roughly 20–200× faster than comparable LLM calls on this class of task
- roughly 40–400× cheaper, with output tokens billed at zero
- “frontier composable intelligence optimized for decisions”
Those figures are TypeSafe’s. The blog is explicit that extraordinary claims need receipts, and it publishes workflow evals, a side-by-side sampling demo, and caveats about how the numbers were produced. The homepage-style “193.6× faster / 444.6× cheaper” comparison is described there as likely near the high end of real-world gains.
The name is a joke with a thesis. Jev is a nod to William Stanley Jevons: make a resource cheaper and people use more of it. TypeSafe is betting that if a decision call is fast and cheap enough to bury inside ordinary software, demand for those calls explodes rather than shrinking the market for intelligence.
Access at launch was a waitlist, with a Discord for builders. The company’s own X account posted a one-line “hello world” the same afternoon.
This digest treats that thread as the canonical public announcement. Later stories on this site cover what builders actually did with the API once they got in.