What is Jev?

Jev is the first public System One model from TypeSafe AI. It is designed for a job that chat models are not optimized for: making bounded decisions that software can use directly.

Instead of asking Jev to write a paragraph, an application sends two things:

1. State — the information the decision should be based on.
2. Questions — the exact choices, scores, or yes/no judgments the application needs.

Jev returns typed answers with probabilities. Your code decides what to do with them.

That makes Jev closer to an intelligent function call than a chatbot. A support system might ask it to choose a queue, score urgency, and estimate whether a case needs a human. The application still owns the workflow, thresholds, permissions, and final action.

What “System One” means

TypeSafe uses “System One” for models built around fast, structured judgment. The name is inspired by the distinction between quick System 1 thinking and slower System 2 reasoning, but the useful difference for a developer is the interface.

A language model normally generates free-form text one token at a time. Even when asked for JSON, an application must parse and validate what it returns. Jev gives up free-form writing and evaluates declared questions in parallel. The possible output shape is fixed before the request is made.

TypeSafe describes the contract as “unstructured state in, typed probabilistic decisions out.” See the official System One overview and the Jev launch post.

The three question types

Jev's interface centers on three primitives:

• Choice selects one option from a set you provide. Use it for classification, routing, or picking an allowed action.
• Score evaluates state against an ordered rubric. Use it for risk, urgency, quality, or fit.
• Noul / Boolean estimates the probability that a statement is true. Use it for gates, guardrails, and review thresholds.

The important constraint is that your application defines the answer space. Jev cannot invent a fourth queue when the request names only three queues. It can still choose the wrong valid option, so probabilities and validation against your own labeled data still matter.

What Jev is good at

Jev is a fit when the next step is a bounded software decision:

• route a ticket, email, tool call, or agent task;
• score risk, urgency, quality, or policy fit;
• verify an output against explicit criteria;
• gate automatic actions and escalate uncertain cases;
• classify large volumes of text into a known taxonomy.

It is not a replacement for a writing assistant, general chat model, or open-ended reasoning system. If the output needs to be an explanation, a draft, code, or a new idea, use a generative model. If the output needs to be one of a few declared decisions, Jev may be the more natural interface.

Speed and pricing claims need context

TypeSafe reports end-to-end latency of 70–500 ms and a direct price of $0.042 per million input tokens with output unmetered in its September 2026 launch material. It also reports large speed and cost advantages over frontier language models on its workflow evaluations.

Those are TypeSafe's own results, not independent guarantees for every workload. The company publishes important caveats alongside them, including geography, workflow design, and benchmark methodology. Read the workflow evaluations before applying the headline numbers to production planning.

Where to try Jev

The most direct path is the TypeSafe quickstart and TypeSafe console. Jev is also listed through OpenRouter and Vercel AI Gateway.

Access terms, model identifiers, and provider pricing can change. Check the provider page you plan to use immediately before implementation.

The short version

Jev is not “a faster ChatGPT.” It is a model for typed decisions inside code. You supply the state and allowed questions; it supplies probabilities; your application remains responsible for the policy.

Next: compare the available routes in How to Access the Jev API.