On Sep 15, 2026, TypeSafe AI announced Jev — the first commercial System One Model: a frontier model that gives up text generation and instead makes fast, calibrated, structured decisions that software can call directly. This independent site tracks the category: what it is, what it's good for, and tools to evaluate it.
Naming borrowed from psychology's System 1 / System 2: LLMs are System 2 — slow, deliberative text generators. A System One Model is the opposite shape: unstructured state in → typed probabilistic decision out, in a single pass, with calibrated probabilities. No prose. No hallucinated strings. Think of it as a frontier-intelligence function call.
TypeSafe reports Jev matches LLM intelligence on decision tasks at ~100× the speed and efficiency. Site will add measured numbers when available.
No string generation means no invented facts. Output is a probability distribution over the options you define.
Trained with Reinforcement Learning for Calibrated Decisions — the numbers it returns are meant to be trusted as numbers.
Routing, moderation, field extraction, risk scoring, A/B variant picking, alert triage — anywhere you currently bolt a JSON-mode prompt onto a chat model and pray.
Model your current LLM function-calling spend vs. a System One approach. Shareable result card.
Pick a scenario, define your options, get the decision schema + curl/JS/Python call snippets.
FAQ, comparison tables, ecosystem tracker (open-source lookalikes, tools), with every claim sourced.
| Item | Status (2026-09-20) |
|---|---|
| Model | Jev (first System One Model), by TypeSafe AI — founded by Diogo Almeida (ex-OpenAI, research behind ChatGPT instruction-following) |
| Announced | Sep 15, 2026 — HN discussion 1900+ points, official post |
| Availability | Early access / waitlist at TypeSafe; also listed on Cloudflare Workers AI (typesafe/jev) |
| Public pricing | Not published yet — our calculator lets you model it with your own assumed price |
| Open-source lookalikes | Early research starters exist (e.g. jevlike, MIT) — toy-scale, NOT Jev-equivalent; see ecosystem notes |