Jev

Diogo Almeida, founder of TypeSafe AI, announced Jev on 15 September 2026 as the first System One Model: text-only, 32k context, and built not to generate text. It takes unstructured input, preferably a program state, and returns typed values whose shape is fixed in advance, with probabilities and confidence scores. Almeida argues that renouncing generation is what rules out hallucination and type errors. Training is RLCD, reinforcement learning for calibrated decisions, and a parallel sampler returns every output in one request. Published figures: $0.042 per million input tokens, free output, 70 to 500 ms latency against 3 to 329 seconds for frontier models. The 193.6× faster, 444.6× cheaper line comes from TypeSafe's own workflow evaluations, and Almeida places it at the high end of real gains.

Assaf Elovic made Jev the default context filter in GPT Researcher, scoring each 1,000-character chunk from 0 to 3 and keeping those at 1.5 or above: 73% relevant passages against 46% for embeddings, 15 wins, 10 ties and 3 losses across 28 tasks, at the same cost. Elovic's reading is that the threshold matters more than the ranking, since Jev without one drops to 50%.

Michelle Chen, Alex Reneau and Kevin Flansburg at Cloudflare cite the buzz around Jev as the starting point for Clef, which they made API-compatible with it, with 64k context and a vision encoder. Clef wins three of TypeSafe's four workflow evaluations and the median latency comparison (209.3 ms against 524.1 ms), while Jev or a variant stays ahead on When2Call, BRIGHT and PhishNChips.