TypeSafe
#16 en Modèles de reasoningTypesafe · 2× · vu le 22 sept. 2026
TypeSafe AI (San Francisco) has released "Jev," the first model in a new category called "System One Models." Unlike classical LLMs, Jev does not generate flowing text, but instead returns structured decisions with calibrated probabilities for pre-defined, typed questions (Choice, Score, Null/Boolean). The model is trained using a proprietary method called "Reinforcement Learning for Calibrated Decisions" (RLCD), which optimizes for reliable, software-compatible confidence values rather than human-preferred text responses. The model is proprietary, offered exclusively through the TypeSafe API/console, and targets production systems that require fast, verifiable individual decisions (e.g., classification, routing, gating).
Fonctionnalités
| Key Benchmark (%) | TypeSafe's own 'JevBench' decision benchmark: 75.3 points (as of Sep 21, 2026, not independently reproduced) |
| Context Window (Tokens) | 32,000 tokens (state + single longest question) or 64,000 tokens (state + all questions combined) |
| License | Proprietary, hosted model, no open-source/on-prem access |
| Multimodality | No – text/JSON input only (state); image, audio, video must be pre-processed into text |
| Platform | Own HTTP API (POST api.typesafe.ai/v1/systemone), Python & JavaScript SDKs; also available via Vercel AI Gateway, Cloudflare, and OpenRouter |
| Price per 1M Tokens | $0.042 per 1M input tokens; output tokens free |
| Release Date | September 15, 2026 (early access / waitlist) |