Getting Started with Jev AI: The Model That Decides Instead of Chatting

Most AI models talk. Ask a question and you get paragraphs back, which is great for a person and annoying for a program. Your code then has to parse the text, hope the model followed your format, and handle the times it didn't.

Jev takes the opposite approach. It doesn't talk at all. You give it a question with fixed possible answers, and it gives you back a decision with probabilities. Think of it as the difference between asking a friend for advice and asking a referee for a call: the referee doesn't explain, they just signal.

What Jev is

Jev is made by TypeSafe AI, a San Francisco start-up founded by Diogo Almeida, previously a researcher at OpenAI. TypeSafe announced Jev on 15 September 2026 as the first of what it calls "System One" models: fast, decisive models built for software rather than conversation.

Jev answers three kinds of question:

Every answer comes with a confidence score, and that's the useful part. "Billing team, 94% sure" and "billing team, 51% sure" should be handled differently by your code.

The numbers TypeSafe claims

• Speed: responses in 70 to 500 milliseconds.
• Price: $0.042 per million input tokens, and output is free.
• Compared with LLMs: up to 193 times faster and 444 times cheaper, according to TypeSafe's own tests.

Treat the multipliers as marketing until you've tested your own workload. TypeSafe ran those comparisons on workflows its own team built. Early independent testers reported accuracy close to large general models on tasks like email sorting and command safety checks, sometimes slightly lower, at a fraction of the cost.

The price is still striking. Classifying a million support messages of about 300 tokens each is roughly 300 million input tokens, about $12.60 at the listed rate.

Getting access

1. Sign up in the TypeSafe console. The waitlist was removed, so early access is open to everyone.
2. Create an API key there and keep it in an environment variable named TYPESAFE_API_KEY.
3. The API lives at https://api.typesafe.ai, with the full request reference at docs.typesafe.ai/api.

Model names follow two patterns: jev-latest always points to the newest stable release, while pinned versions like jev-1.13.0 never change. Pin a version in production, especially if your code acts on confidence thresholds, so an update doesn't quietly shift your results.

Your first question

The quickest way to try Jev without writing an API client is the open-source evaluation tool promptfoo, which has a TypeSafe provider. Questions are defined by ID, with a type and instructions:

questions:
  team:
    type: choice
    instructions: Which team should handle this ticket?
    criteria:
      billing: Payments, invoicing, refunds
  urgent:
    type: noul
    instructions: Does the customer need a response today?

Each input you test is sent as the state, and every question is answered against that same state. For a ticket like "I was charged twice for my October invoice, please fix before my card limit runs out", you'd expect billing with high probability for the first question and a high "yes" for urgency.

If your state is JSON, Jev receives it as structured data, so a question can refer to fields by name: "Is order_total unusually high for this customer?"

Where Jev shines

• Routing: send each ticket, lead or email to the right team.
• Moderation: flag insults, spam or off-topic posts, with a confidence you can threshold.
• Triage and scoring: rate urgency 1 to 5, or sentiment on a fixed scale.
• Agent decisions: let a fast model make the hundreds of small "which tool next?" choices inside an AI agent, and save the expensive model for the parts that need writing or reasoning.
• Evaluation: grade outputs from other models against a rubric, cheaply.

Where it doesn't

• It can't chat, explain itself or write a reply. You'll still need a general model to draft the actual answer to that billing ticket.
• It only picks from your options. If none of them fit, it can't say so unless you include an option like "none of these".
• It works best when each question asks one precise thing. "Is this urgent and angry and about billing?" should be three questions.
• Requests are limited to about 64,000 tokens, with up to 32,000 for the state, and inputs are text only (strings or JSON).

A sensible way to start

Pick one boring, high-volume decision your team or app makes every day, like tagging inbound emails. Run Jev and your current approach side by side on a few hundred real examples. Compare accuracy, look at what happens below 70% confidence, and decide where a human should step in. If the numbers hold up, you've just turned an expensive step into a nearly free one.

In one line: Jev answers closed questions with probabilities instead of text, which makes it fast and cheap for decisions inside software, and useless for anything that needs writing.