Diogo Almeida, a former OpenAI researcher who helped create ChatGPT and develop reinforcement learning from human feedback, has released Jev, a new type of transformer-based AI model through his startup TypeSafe AI. Unlike conventional large language models, Jev does not output text but instead produces calibrated probability scores, making it dramatically cheaper, faster, and incapable of hallucination. The model has generated considerable developer interest, with the company's API briefly unable to handle the surge in demand.

Jev represents a fundamental rethink of AI model design for automation tasks. In testing, Vercel achieved 5 to 18 times faster results with greater accuracy than OpenAI's Luna model for safety command classification, whilst Bryo AI found Jev 10 to 20 times less expensive than Gemini for email classification. The model's free output tokens and billion-metered input pricing make it well-suited for software automation, model routing, and serving as a verification mechanism for other AI agents, allowing users to evaluate outputs based on confidence scores rather than accepting decisions blindly.