Jev: A Classification Model That Does Not Hallucinate

Diogo Almeida, one of the creators of ChatGPT and the author of the RLHF method, has released Jev through his startup TypeSafe AI. The model outputs probabilities across predefined categories instead of text, and therefore does not hallucinate. In Vercel's tests, Jev runs 5 to 18 times faster than ChatGPT Luna 5.6 on classification tasks. According to Bryo AI, it is 10 to 20 times cheaper than Gemini on email parsing, though slightly less accurate.

For agent pipelines, Jev fits cases where only a quick signal is needed (pass or stop) rather than a full answer. I would place Jev as a filter before expensive Opus calls. It handles simple steps like command classification at minimal cost, and leaves the heavy work to the frontier model. The overall run cost drops.