Release2026-09-14
TypeSafe released its first System One model, Jev. It outputs type-safe structured values instead of text. It runs 40x–200x faster than frontier LLMs on System One tasks. Output tokens are free. Input tokens cost $0.042 per million. It never makes type errors. It always returns calibrated probabilities. It is optimized for verifiable, structured decision tasks. Early access is open now. The model uses a new training method called RLCD. It samples outputs in parallel rather than token-by-token.
What changed
Frontier models generated text sequentially, with variable confidence and type errors.
What it unlocks
Running AI as a fast, type-safe function call inside software with calibrated probabilities.
- $0.042 per million input tokens
- 70ms–500ms response time
- 193.6x faster, 444.6x cheaper on workflow evals
- 255 output cardinality limit
What you need to act on it
- early access to TypeSafe
- structured program state as input
Sources