What Is Jev AI? A Complete Guide

Jev is a decision model from TypeSafe AI, the first of what the company calls System One models. Where a large language model generates text, Jev reads a situation and returns a structured decision: a category, a score, or a true/false judgment, each with calibrated probabilities. It is fast and cheap, and it cannot write a sentence.

If you have ever spent tokens and a few seconds on an LLM call just to get back one word like billing, Jev is aimed at that problem.

What it does

The tasks are narrow but extremely common:

• Classification — label support tickets, emails, reviews, or user intents
• Routing — decide which agent, department, queue, or model should handle an input
• Tool selection — pick the next tool or sub-agent in an automated workflow
• Scoring — rate leads, answers, content quality, or risk on a defined rubric
• Validation and guardrails — check whether a statement or output is acceptable before it moves on

All of them share the same shape: structured input in, structured decision out, with a confidence value your code can act on.

How it differs from an LLM

Models like GPT and Claude are general-purpose generators. That generality is the point of them, and it is also why they are an awkward fit for decisions.

The pattern that works is not "Jev instead of an LLM." It is Jev in front of them: Jev handles routing and gating, and the LLM is reserved for the work that actually needs generation.

State and questions

You send Jev a state (the text or context to judge) and any number of typed questions. There are three question types: Choice (pick from described options), Score (place on a rubric), and Noul (is this statement true, 0–1). Every question is evaluated against the same state in parallel and in isolation, so adding questions barely changes response time.

A minimal exchange sends a customer message as the state with a Choice question such as "which team should handle this?", and gets back the selected option, a probability for every option, and a confidence value. No prompt engineering, no parsing free text. The API guide shows the exact requests and responses.

Where you can use it

Jev is served through the POST /v1/systemone HTTP API with official Python and JavaScript SDKs, and it is available through the Vercel AI Gateway. If you already route AI traffic there, adding Jev is a configuration change rather than a new vendor integration. There is also an official playground where you can paste text and see decisions immediately.

One thing to know before you commit: as of Jev 1.13 the input is text-only, with no images, audio, or video. English accuracy is the strongest, other languages work but unevenly, and Jev is not trained on your customer data.

Where Jev came from

Jev is built by TypeSafe AI, a San Francisco lab founded by Diogo Almeida, Sasha Sheng, and Erik Gafni. Almeida co-invented RLHF and InstructGPT, the research behind ChatGPT and GPT-4, and worked before that at OpenAI and Google Brain. Sheng came from Meta/FAIR. Gafni is a repeat founder. The team spent roughly two years in stealth before announcing Jev on September 15, 2026.

The launch landed harder than you would expect for a model that cannot write a sentence. The Hacker News thread reached 1,949 points and 511 comments, and inside a week there were community ports, local reimplementations, CLI wrappers, and guardrail tools built on top of it. TypeSafe calls Jev a System One model, borrowing Daniel Kahneman's name for fast, intuitive thinking. The model itself is named after the economist William Stanley Jevons.

How good is it, really?

TypeSafe publishes its own workflow evaluations. They are vendor-run, so read them the way you would read any vendor's numbers, but they are unusually detailed. Across four production-shaped workflows (security incidents, agent-trace review, invoice processing, customer service), Jev averages 67.8% accuracy at $0.0004 and 0.4 seconds per case. GPT-5.6 Terra averages 67.9% on the same set. Claude Opus 5 averages 73.1%.

So Jev is not more accurate. It is about as accurate for a small fraction of the cost and time. TypeSafe's headline from those workflows is 193.6x faster and 444.6x cheaper. The full queries and methodology are published at evals.typesafe.ai if you want to pull the numbers apart yourself.

When it is the right tool

Jev fits when you call a model on every request and cost or latency matters, when the answer space is known up front, when you need a calibrated confidence value to drive fallback logic, and when you want typed, consistent decisions rather than creative variation.

It is the wrong tool for long-form generation, multi-step reasoning over complex documents, or open-ended creativity. Keep an LLM for those.

Try it

Reading about decision models only gets you so far. The fastest way to understand Jev is to feed it a real decision from your own product and watch the probabilities and confidence come back.

You can practice routing, tool selection, and evaluation scenarios in the Jev Agent playground.

What to read next

• Jev Model Explained: System One Models
• Jev vs LLM: What's the Difference?
• Jev Tutorial: Build Your First Decision Workflow