Jev, a model that answers with probabilities instead of text
TypeSafe AI's Jev takes a block of state and typed questions (yes/no, a choice among up to 255 options, a score on a rubric) and returns calibrated probabilities in about 100 ms, at $0.042 per million input tokens with output free. Two weeks after launch, independent runs put it a few points behind Opus on classification at a twentieth of the latency.
Jev launched on 15 September as the first "System One" model from TypeSafe AI, a company that came out of stealth the same day with a reported $40 million seed round. It does not write text: a request carries a state (a document, a record, a JSON object) and a set of typed questions evaluated together, and each answer is a probability per option, a yes/no probability, or a probability-weighted score, trained for calibration with what TypeSafe calls reinforcement learning for calibrated decisions. The vendor quotes about 100 ms per query, $0.042 per million input tokens and free output, up to 255 options per choice; it is hosted only, with Python and JavaScript SDKs, and is not trained on customer requests. Vercel, Cloudflare, OpenRouter and Databricks added access within ten days, and signups were paused on 22 September for demand. The independent evidence is now arriving: OpenRouter's Banking77 run (3,080 utterances, 77 intents) scored Jev 1.13 at 81.0% against Claude Opus 5 at 84.4%, at a median of 175 ms against 2,266 ms and $0.11 against $2.42 per thousand requests, while finding its confidence score ranks well but is not a calibrated probability; a 24 September arXiv study of 2,170 public Jev projects found attribute judgment and scoring the dominant uses; a 30 September study of recommendation reranking found it effective with gentler latency growth than LLM rerankers, but slower than purpose-built recommenders. The fit is narrow and real: a fixed option set, one decision per request, and a probability the surrounding code can threshold.
Read the source