What is a “System One model”?

An LLM answers a question by generating text, token by token, autoregressively - which is flexible but slow and expensive for anything that’s really a structured decision: is this spam, which category does this fall into, does this answer agree with that one. A System One model (the name is a nod to Kahneman’s fast, intuitive “System 1” thinking) skips generation entirely: you send it a state and a set of typed questions, and it returns typed answers with a calibrated confidence in a single parallel pass.

TypeSafe AI’s Jev is the first product in this category, launched September 2026. Independent testing since launch has found its speed and cost claims hold up, and - the part that actually matters for using it to gate decisions - that its confidence scores are genuinely calibrated rather than just confident-sounding. See how we use it.

Why this matters for routing and verification

A router or a verifier answers structured questions all day: which model tier, does this need escalating, do these two answers agree. That’s exactly the shape a System One model is built for, at a fraction of the latency and cost of asking an LLM to answer the same question in free text and then parsing it back out.