01What Is Jev AI and How It Works

Jev AI is a new type of AI model that makes decisions, not conversations. It was released by TypeSafe AI, and the person behind it is Diogo Almeida, one of the co-inventors of Chat Cheaper TA. He spent two years building a new way to train models.

Jev does not write emails, code, or explanations. You give it a question with set answers, and it picks one.

Jev is called a "system one" model because it works like the fast part of your brain. When you see a red light, you stop without writing a paragraph about it. Jev does that for your AI agents.

It handles tiny decisions that used to go to big, slow models. This makes agents 20 to 200 times faster and 40 to 400 times cheaper.

You hand Jev the current situation and a list of options. It returns the choice, a probability for each option, and an overall confidence score. This lets you automate the boring decisions while keeping control when it matters.

02Main Features of Jev AI

Jev AI has three types of questions you can ask, and they cover almost every decision:

  • Choice – Pick from a list. Example: "Which step goes next?" It gives the pick plus a probability for every option.
  • Score – Rate something on a scale you define. Example: "How relevant is this source?" It gives a number that can land between levels.
  • Null – Yes or no. Example: "Does this need human approval?" It gives the probability, like 0.999 means almost certainly yes.

You can ask many questions in one request, and they get answered in parallel. Asking five questions at once barely changes the speed or cost. This is huge because it removes the waiting time in your agents.

Another key feature is confidence. Jev tells you how sure it is. You can set a threshold: above it, automatically continue; below it, stop and ask a human.

03How to Set Up Jev AI in Your Agents

Getting started with Jev AI is straightforward, even if you are not a developer. Follow these steps:

  1. Find Jev on OpenRouter, where it is currently in beta. You can access it like any other model.
  2. Define the situation you want to make a decision about. Be specific, like "This is an email from a client asking for a refund."
  3. List your options with clear descriptions. For example: "Reply immediately," "Research first," or "Flag for human review."
  4. Send the question and options to Jev. It will return the choice and a confidence score.
  5. Set a confidence threshold. If Jev is above it, let your agent act automatically. If below, route it to a human.
  6. Measure the cost. Jev charges about four cents per million tokens sent in. No charge for what comes back.

Start with a simple decision you make repeatedly. Then, expand to more complex ones as you get comfortable.

04How to Use Jev AI with Examples

People built impressive things with Jev in the first few days. Here are three examples from the episode:

Research sorting: A developer named Hassan summarized 1,018 AI research papers, then sent each title and summary to Jev with 24 possible topics. Jev picked the category for all papers at a cost of eight cents total. The median time per paper was 256 milliseconds.

Email sorting: Riley Brown ran 500 emails through Jev. It sorted them in seconds for three and a half cents per email. Options included "reply," "research first," and "flag for me." Each answer pointed the email to a different folder or helper.

Lead scoring: A builder named Roman fed Jev 700 leads with personalized outreach messages. Jev scored each one in 40 seconds for nine cents, predicting how each message would land and flagging mismatches.

These show how Jev handles the small decisions that used to slow down agents, letting you focus on the big picture.

05Results and Comparisons from Tests

The episode showed several tests with clear numbers:

  • Context compression: Plugging Jev into Claude Code cut a session from nearly 1 million tokens to 86,000 in one second. People reported context drops of 30-60%, with one case going from 90% down to 9%.
  • Browser agent: The browser-use team put Jev inside a browser agent. It found flights in seven seconds for under half a cent. Their tuned version cut the median browser commands from 1,092 to 101, and median task time dropped 25%.
  • Internal linking: A person named Borgia pointed Jev at his 586-page website to rebuild internal links. Jev did it in 45.1 seconds, placed 584 links, and left 139 pages unlinked because nothing fit. Running the same job with Claude Opus 5, it only got through 21 pages. That's 586 vs. 21.

These results show Jev is not just faster; it also decides when to do nothing, which prevents forced links and errors.

06Tips, Limits, and Mistakes to Avoid

Jev AI is powerful, but it has limits. Here are tips and warnings from the episode:

  • It can't write anything. Research, drafts, and files still need normal models. Jev only picks.
  • Confidence is not accuracy. A high confidence score means the model is sure, not that it is correct. Always verify important outcomes.
  • It can be tricked. If someone feeds malicious text into the situation, they can push the decision. Be careful what you put in.
  • Field names don't matter. Jev does not see the name of the question. You must put the instruction in the question itself and describe options clearly.
  • Provide proper context. Saying "the researcher finished" gives Jev nothing. Give it sources, findings, and what is missing.
  • Beware of cheap decisions. A cheap decision that sends your agent down the wrong path can cost more than a safe one. Watch cost per finished task, not cost per decision.

Use Jev for repeated admin tasks like sorting, scoring, routing, and flagging. That is where the real gains are.

07Who Should Use Jev AI and Why It Matters

Jev AI matters most to business owners, not just developers. The biggest barrier to handing work to AI agents is trust, not capability. You know the agent can write an email, but you don't know if it will send the right one.

Jev solves that by giving you a confidence number. If it says 62% sure, you can look at it before it goes.

The episode showed that the jobs Jev does are admin tasks: sorting 500 emails, scoring 700 leads, deciding which page links to which, routing work to helpers. These are tasks you do manually or pay someone else to do. Now they can be automated cheaply and fast.

The gap between people who set this up and those who just read about it will widen quickly. You don't need to be technical. Every result came from someone describing a decision in plain English and giving the model a list of options.