What Is Jev? The AI That Decides Instead of Writing

Most people use AI the same way for everything: open a chat, paste something in, get a paragraph back. That works for writing. It's a waste for the hundreds of small decisions that make up a working day.

Is this email urgent? Is this lead a real buyer? Which team should handle this request? Is this customer asking for a refund? Those aren't writing tasks. They're sorting tasks. And until recently we've been paying a writer to do a sorter's job.

That's the gap Jev fills.

What is Jev, exactly?

Jev is an AI model built by TypeSafe, a lab founded by Diogo Almeida, Erik Gafni and Sasha Sheng. It went into early access on 15 September 2026 and is available through OpenRouter. The name isn't an acronym.

A normal language model (ChatGPT, Claude, Gemini) predicts words. Jev doesn't produce words at all. You describe a decision, and it hands back a structured answer in one of three shapes:

You explain each option in plain language, the way you'd brief a new hire. Jev reads those descriptions, reads your text, and decides. The quality of its answer depends on how clearly you describe the options. That's the same lesson as with any AI: vague instructions get the most probable answer, not the one you meant.

Jev vs ChatGPT or Claude: what's the difference?

They're not competitors. They do different jobs.

Think of it as hiring two people. The chatbot is your writer. Jev is your sorter: fast, cheap, and it only ever ticks one of the boxes you gave it.

How much does Jev cost?

This is the part that changes what's worth automating. On OpenRouter, Jev costs about $0.04 per million input tokens, and output is free. OpenRouter's own example works out to roughly $19 for a million support tickets.

In a How I AI episode on Lenny's Newsletter, Claire Vo (founder of ChatPRD) used it to categorise 1,700 pull requests for about 9 cents, and to turn 4,500 YouTube comments into an audience-insights dashboard.

At those prices, jobs you'd never pay a chatbot to do one by one become reasonable: scoring every email that hits your inbox, tagging every lead in your CRM, reading every review a competitor ever received.

What can a business actually use Jev for?

Here's how it maps to the kind of work my coaching clients bring me:

• Inbox triage. Score every incoming email for urgency and route it: "client issue", "new lead", "supplier", "ignore". A chatbot then drafts replies only for the ones that matter.
• Lead qualification. Is this enquiry a real buyer, a tyre-kicker or a supplier pitching you? Act automatically on the clear ones, look yourself at the unsure ones.
• Support ticket routing. Billing, technical or account? Is the customer asking for a refund? Is it urgent?
• Feedback at scale. Tag hundreds of reviews, survey answers or social comments by theme and sentiment, then let a chatbot write the summary.
• Checking an AI agent before it acts. Before an agent sends an email or deletes a file, ask Jev: "Does this follow our policy?" Low confidence means a human looks first.

Why the confidence number matters

Jev doesn't just answer. It tells you how sure it is, and TypeSafe calibrated those numbers so that, across many decisions, "80% sure" really means right about 80% of the time. (Any single answer can still be wrong.)

That lets you set three simple bands instead of trusting everything or nothing:

1. High confidence: act automatically.
2. Middle: flag it for review.
3. Low confidence: send it to a human.

Where to draw those lines depends on your business. Test it on 50–100 real examples you've already sorted by hand, and see where it starts making mistakes.

The catch: use it with a chatbot, not instead of one

Jev has clear limits:

• Text only. No images, PDFs as scans, audio or video.
• No explanations. You get a number, not the reasoning behind it.
• Not for maths or rules. Dates, totals and thresholds belong in a spreadsheet or a simple script, not in any AI model.
• Hosted and proprietary. You depend on TypeSafe's service, and answers can shift slightly between calls.

Claire Vo said she stopped using Jev on its own. The pattern that works is a pair: Jev decides, a chatbot writes. Jev scores and routes the emails; Claude drafts replies for the five that matter.

If that sounds like a workflow with separate jobs and a checker, it is. It's exactly the idea behind graph engineering: split the work into jobs, give each job the right tool, and keep a human where mistakes get expensive. Jev is a natural fit for the "sorter" and "checker" jobs in that graph.

How to start with Jev if you don't code

Jev is used through an API (OpenRouter or TypeSafe's own SDKs), so there's no chat window to type into. That sounds like a developer-only tool. In 2026 it isn't, because you can ask a coding agent such as Claude Code to wire it up for you.

Step 1: Pick one sorting decision you make every day

One, not five. "Is this email from a potential client?" is a great first one.

Step 2: Write the options like a brief for a new hire

For each option, two or three sentences on what counts and what doesn't. This is 80% of the result.

Step 3: Collect 50 real examples you've already sorted

This is your test set. Without it you can't tell whether the model is good or just confident.

Step 4: Ask a coding agent to build the test

I want to test the Jev decision model (model "typesafe/jev-1.13" on OpenRouter) on one decision in my business.

The decision: [e.g. "Is this incoming email from a potential client?"]
The options and what each means: [paste your brief]
My test set: [attach a CSV with the text and the answer I chose by hand]

Build a small script that sends each example to Jev, saves its answer and confidence next to mine, and shows me:
1. How often Jev agreed with me
2. Accuracy above and below different confidence levels, so I can pick my "act / review / escalate" thresholds
3. The 10 examples where it was most confidently wrong

Ask me for my OpenRouter API key at runtime. Never write it into a file.

Step 5: Only then connect it to your inbox or CRM

Once the numbers hold up on your own examples, automate it. Start with "flag for review", not "act automatically".

Is Jev worth it for you?

If most of your AI use is writing, Jev changes nothing. Keep using your chatbot.

If your day is full of small, repeated judgments (sorting leads, triaging messages, tagging feedback, checking work before it goes out), Jev is the first tool that makes automating them almost free. The skill that matters isn't the API. It's being able to describe a decision clearly enough that a machine can make it. That's a skill worth building whatever the next model is called.

FAQ

Sources: OpenRouter: What is Jev (pricing, question types, calibration, limits) · Lenny's Newsletter / How I AI: Jev for beginners with Claire Vo (use cases and costs). Prices and model version as of October 2026.