TypeSafe AI Emerges From Stealth to Launch Jev, the First System One Model
After two years in stealth, former OpenAI researcher Diogo Almeida and TypeSafe AI launched Jev—a non-generative model returning typed probabilities in 70–500 ms at $0.042/M tokens.
After two years operating in stealth with a reported $40 million seed round led by DCVC, San Francisco startup TypeSafe AI launched publicly on September 15, 2026, introducing Jev—the first commercial “System One model” built to return fast, type-safe probabilistic decisions rather than conversational text.
A frontier-intelligence function call for software
Founded by former OpenAI researcher Diogo Almeida alongside Sasha Sheng and Erik Gafni, TypeSafe built Jev on a dedicated stack combining a specialized transformer architecture, a parallel sampler that answers multiple questions in one forward pass, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD).
| Dimension | Conversational LLMs | TypeSafe System One + Jev |
|---|---|---|
| Training objective | RLHF / RLVR | RLCD (Calibrated Decisions) |
| Outputs | Free-form text strings (parse + validate) | Typed Choice, Score, Noul probabilities |
| Sampling | Sequential token-by-token | Parallel single-pass evaluation |
| Pricing | $0.20–$10/M input; ~5× output | $0.042/M input; output free |
| Latency | 3 to 329 s (frontier reasoning) | 70–500 ms end to end |
“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
Diogo Almeida — Introducing System One Models & Jev (Sep 15, 2026)
Coming just four days after TypeSafe published its September 11 anti-benchmaxxing manifesto (“Lies, Damned Lies, and Benchmarks”), the launch immediately reshaped how AI engineers and major cloud labs approached software automation.