TypeSafe AI Introduces System One Models and Jev

TypeSafe AI announced Jev, its first System One Model, on September 15, 2026. Unlike a chat LLM, Jev doesn’t generate strings — it takes unstructured state as input and returns typed, probabilistic decisions in a single parallel pass instead of token-by-token generation.

The headline numbers from the announcement: end-to-end response time of 70-500ms (versus 3-329 seconds for frontier LLMs on comparable tasks), input pricing of $0.042 per million tokens with output free, and — because outputs are structured by construction rather than parsed out of generated text — no type errors are possible.

Jev is trained with a new method TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for epistemically honest probabilities on a decision rather than for the kind of response a human rater would prefer.

Early access is open now, with TypeSafe bringing developers off the waitlist as capacity allows.