Typesafe AI introduces jev, a high-performance decision-making model

TypeSafe AI has launched Jev, its first System One model, designed for making typed decisions with calibrated probabilities. Unlike traditional models that chat, write code, or summarize, Jev processes unstructured state and returns structured decisions. This makes it ideal for use within agent loops, where it can assess safety, relevance, and completion status. The model is trained using Reinforcement Learning for Calibrated Decisions (RLCD), ensuring higher confidence correlates with higher accuracy. TypeSafe claims Jev is 193.6x faster and 444.6x cheaper than existing solutions, based on internal evaluations. In a head-to-head test with Claude Opus 5, Jev was 3.3 points less accurate but 13x faster and about 1/22 the cost. The model supports up to 255 decision options and operates by evaluating multiple typed questions in parallel against the same state. Open-model performance figures are self-reported, but OpenRouter’s Banking77 test provides a clear benchmark for comparison.