TypeSafe AI

Commercial provider · 2 models tracked

A San Francisco lab that deliberately took the opposite research direction to chat: instead of RLHF, it trains with what it calls Reinforcement Learning for Calibrated Decisions (RLCD) to build "System One models" — models meant to be called by software rather than read by people. They evaluate a block of state against typed questions and return decisions with calibrated probabilities; they do not write replies, produce code, or explain their reasoning. It sells two versions of one model, Jev 1.13 and Jev 1.14 (released October 14, 2026), on one endpoint, and bills input tokens only at $0.042 per 1M for both, with output tokens free. TypeSafe states no founding year on any of its own pages, so none is published here.