Built by a researcher who helped invent RLHF, Jev outputs probabilities instead of words, promising cheaper, faster and hallucination-free automation.
TechCrunch profiles Jev, a new model from TypeSafe AI, founded by Diogo Almeida, a former OpenAI researcher who helped develop reinforcement learning from human feedback. Instead of generating text, Jev outputs probabilities and "calibrated decisions" for a set of options defined in advance, which the company says eliminates hallucinations because the possible outputs are fixed up front.
The economics are the headline. Output tokens are free and input is metered by the billion rather than the million, and developers report large gains in practice: Vercel measured five-to-eighteen-times speedups, while Bryo AI found Jev roughly ten-to-twenty-times cheaper than Gemini for email classification. The model is trained exclusively on synthetic data using a method Almeida calls "reinforcement learning from calibrated decisions," and is named after the economist William Stanley Jevons.
Why it matters: deterministic, low-cost models for classification and routing map cleanly onto enterprise workflows, where reliability and cost discipline often matter more than open-ended generation, a pattern we build into agentic systems.