Typesafe Ai · depuis 15. September 2026 (Early Access) · 9× · vu le 19 sept. 2026
Jev is TypeSafe AI's first "System One Model," released in early access (waitlist) in September 2026. Unlike conventional LLMs, Jev does not generate free text; instead, it evaluates predefined typed questions (Choice, Score, Noul/yes-no) and returns structured, calibrated probability values in a single parallel inference pass rather than token-by-token generation. It is trained with a proprietary method called "Reinforcement Learning for Calibrated Decisions" (RLCD) and is available only via a hosted API, with no open weights or self-hosting option. According to TypeSafe's own "workflow evals," Jev achieves decision accuracy comparable to mid-tier LLMs at much lower cost and latency, though it trails top frontier models like GPT-5.6 Sol or Claude Opus 5 on raw accuracy.
Fonctionnalités
| Key Benchmark (%) | 67.8% mean accuracy on TypeSafe's own workflow eval (4 tasks, 711 cases), vs. 74.1% for the best comparator model |
| Context Window (Tokens) | approx. 32,000 tokens (~150,000 characters) |
| License | Proprietary, closed-source, not an open-weights model |
| Multimodality | No image or audio input; text/structured state input only, no free-text output |
| Platform | Hosted API only (POST api.typesafe.ai/v1/systemone), model ID jev-latest; Python/JavaScript SDKs; no self-hosting |
| Price per 1M Tokens | $0.042 per 1M input tokens; output tokens free |
| Release Date | September 15, 2026 (early access, waitlist) |