Jev is a System One model from TypeSafe AI — an 🏷️#ai-model that returns typed, probabilistic decisions instead of generated text. Released September 15, 2026 in early access, it takes a block of program state plus a set of typed questions and evaluates them in a single parallel pass, returning answers with calibrated confidence scores. Pricing is $0.042 per million input tokens with output tokens free, billed per input token through a usage-metered API rather than a subscription.
TypeSafe AI was founded by Diogo Almeida, a former 📝OpenAI researcher and co-inventor of RLHF, and spent two years in stealth before launching Jev with $40M led by DCVC. The company's argument is that RLHF-trained models were optimized to please humans rather than to decide reliably, and that software making high-volume repeated decisions needs a different primitive — one trained through Reinforcement Learning for Calibrated Decisions (RLCD) instead of next-token prediction. TypeSafe reports 40–200x faster inference and 40–400x lower cost than comparable LLMs on classification work, at roughly 68% accuracy on its own four-workflow benchmark. Official SDKs cover 📝JavaScript (@typesafe-ai/sdk, Node 20+) and Python, with a 📝Langchain integration exposing it as TypeSafeClassifier.
Key Features
- Choice — Returns one option from a typed union you define, arriving as a narrowed 📝TypeScript key your code branches on without parsing.
- Score — Positions the state on a rubric you define, returning a numeric value between your labeled anchors rather than a bucket name.
- Noul — Answers a yes/no question as a probability from 0 to 1, leaving every threshold for acting or escalating to your own code.
- Parallel evaluation — Ingests the state once and answers every question in the request simultaneously, so more questions cost tokens but almost no latency.
- Calibrated confidence — Each answer carries a confidence estimate, calibrated across groups of predictions rather than guaranteeing any individual answer.