Chat models got incredibly good at sounding helpful. That did not magically produce reliable automation.
In September 2026, TypeSafe AI came out of stealth with a different bet: System One models — models built to make fast, structured decisions that software can consume directly. Their first public model is Jev.
If you’ve ever asked an LLM to “return JSON” and then spent an afternoon parsing, validating, and babysitting edge cases, Jev is aiming at that exact pain.
What Jev actually is
Jev is not a smaller ChatGPT. It does not generate free-form text. You give it:
- State — the thing to judge (a support ticket, a product listing, a JSON blob of program state, a Wikipedia page, etc.)
- Questions — typed questions with fixed answer shapes
It returns typed answers with probabilities (and, for some question types, confidence scores). TypeSafe’s framing: think of Jev as a frontier-intelligence function call — unstructured state in, typed probabilistic decisions out.
That design choice buys three things that chat models struggle with when they sit inside a loop:
- No type errors by construction. Possible outputs are defined up front.
- Parallel evaluation. Multiple questions over the same state can be answered in one request.
- Speed and cost tuned for software, not essays. TypeSafe reports end-to-end responses in roughly 70–500ms, with input priced around $0.042 per million tokens and output tokens treated as too cheap to meter.
Details live in the announcement and the docs intro.
The three question primitives
TypeSafe exposes three question types you can mix in a single call:
| Type | Question shape | What you get back |
|---|---|---|
| Noul | Is this statement true? | A probability from 0–1 |
| Choice | Which option from a closed set? | Selected option + per-option probabilities + confidence |
| Score | Where does this sit on an ordered rubric? | Score + level probabilities + confidence |
A support ticket might ask, in one round trip: Is this urgent? (noul), Which team owns it? (choice), and How frustrated is the customer? (score). Your code then routes.
Where the name comes from
“System One” nods to Kahneman’s fast/intuitive System 1 thinking — except TypeSafe’s pitch is that calibrated, constrained System One models can be more reliable for automation than free-form System 2 chat.
“Jev” is named after William Stanley Jevons. TypeSafe’s joke-with-a-thesis: when intelligence gets dramatically cheaper, demand for intelligence rises — the Jevons paradox applied to AI.
What this is not
Jev will not write your README, invent a refund email, or brainstorm product names. Generation is the wrong tool for the wrong job. TypeSafe’s own jaggedness notes for jev-1.13 are refreshingly blunt: keep math in code, avoid counting with the model, don’t ask it to generate text by chaining choices.
How to try it
- Docs and quickstart: docs.typesafe.ai
- HTTP API:
POST https://api.typesafe.ai/v1/systemone - Python SDK:
typesafe-sdk· JS SDK:@typesafe-ai/sdk - Also available via gateways such as Vercel AI Gateway (
typesafe-ai/jev) and Netlify AI Gateway
Early access is rolling. If your product is full of “if this looks like X, do Y” decisions that you’ve been papering over with prompts, Jev is the first mainstream model that treats those decisions as the product — not as an afterthought glued onto a chatbot.
Comments
Loading comments…