01What is Jev AI and How Does It Work?
Jev AI is a new type of AI model that makes decisions instead of writing. You give it a situation, a question, and a list of allowed answers. It picks one and tells you how sure it is.
It never writes emails, code, or explanations.
Jev AI is built by TypeSafe AI, started by Diogo Almeida, one of the co-inventors of ChatGPT. It is called a system one model, like the fast part of your brain when you see a red light and just stop. This makes it 20 to 200 times faster and 40 to 400 times cheaper than normal writing models.
There are three question types you can ask: choice (pick from a list), score (rate against your levels), and null (yes/no with a probability). You can ask many questions at once, and they all come back together fast.
02Key Features: Routing, Safety, and Confidence
Two features make Jev AI useful for developers and business owners. First, it can route between models. You describe what each model is good at, and Jev picks the cheapest one that can finish the job.
LangChain shipped an integration for this, so you can see the cost savings live.
Second, it can safety-check tool use. You can wrap it around your AI agent to watch tool calls and block risky actions before they run. This guardrail used to be locked inside other products, but now you can add it to your own setup.
It also gives a confidence score for every answer. You set a threshold: above it, your agent acts automatically; below it, it asks you. This helps you trust automation because you know when the AI is unsure.
You can also batch multiple questions in one request, barely changing speed or cost.
03How to Get Started with Jev AI for Free
Getting started with Jev AI is simple, and you can use it for free until September 25 (as of making this episode). After that, you can get it directly from OpenRouter.
- Go to OpenRouter and create an API key.
- Use that key in any AI agent like ChatGPT or Claude Code.
- Add Jev AI as the model for decision-making tasks.
- For the voice browser example, get the GitHub project called 'Jev voice browser' by Moritz Kremb.
- Install it with four lines in a terminal (check the code first).
- Use your OpenRouter key with it, then access it via a local page in Chrome.
04How to Use Jev AI: Real-World Examples
In this episode, Julian shows several ways to use Jev AI. One is a voice-controlled browser. You say commands like 'Open Google' or 'Scroll down,' and Jev makes decisions about what to click or type.
It asks about nine questions per request and answers in about 400 milliseconds.
Another example is a real-time outfit picker app. Julian built an app where you speak your destination (like 'I'm giving a keynote'), and Jev picks a jacket, shirt, and trousers from a preset closet. Each item has a one-line description, and Jev uses that to choose.
The app runs in real time because Jev answers in under a second.
- Email sorting: Jev sorts 500 emails in seconds for 3.5 cents.
- Lead scoring: Jev scores 700 leads with personalized messages for 9 cents.
- Internal linking: Jev rebuilt the link map for 586 pages in 45.1 seconds.
You can also integrate Jev with coding agents like Claude Code. You tell it what to do in plain English, and it writes the app while Jev handles all the small decisions.
05Results and Tests: Speed and Cost Savings
In this episode, Julian shares specific tests and results. Developers ran 1,018 research papers through Jev to categorize them; it cost $0.08 total, with a median time of 256 milliseconds per paper. Riley Brown sorted 500 emails for 3.5 cents.
Roman scored 700 leads for 9 cents.
Borgia rebuilt internal links for 586 pages in 45.1 seconds for 21 cents. The same job with Claude Opus 5 only got through 21 pages. Jev placed 584 links and left 139 pages unlinked because nothing fit, which is important to avoid forcing bad links.
For the voice browser, Julian tested eight commands. Jev answered nine questions per request with a median time of 445 milliseconds and got all eight right. Claude Haiku took 1.6 seconds, Gemini 2.2, and GPT-5 Luna 2.5, each getting seven out of eight.
In another test, Jev answered in 413 milliseconds, while Claude Haiku took 1,486 ms, Gemini 3.2 seconds, and Claude Sonnet 5 4.5 seconds.
A context compression test showed a session dropping from nearly 1 million tokens to 86,000 in one second.
06Tips, Limits, and Mistakes to Avoid
Jev AI can't write anything. It only picks from your options. If you need an email drafted, use a normal model for that.
Jev is for decisions only.
Confidence is not accuracy. A high confidence score means the model is sure, but it could still be wrong. Always check important outcomes separately, like the browser use team does after Jev says 'done.'
Jev can be tricked by text in the situation you give it, so be careful what you include. Also, Jev doesn't see the name of your question field; the instruction must be in the question itself, and options must be described properly.
- Jev charges for what you send in, not for what comes back. Input cost is about $0.042 per million tokens.
- You can set confidence thresholds: above 0.45, act; below, ask the user.
- Jev can only pick from the list you give it, so always include a 'do nothing' or 'keep current' option.
- For long tasks, be cautious about deleting history; it can lose the reasoning trail.
07Who Should Use Jev AI and Why It Matters
Jev AI is for anyone who makes repetitive decisions, not just developers. Sorting emails, scoring leads, routing tasks, and deciding which page links to which are admin tasks you can now automate cheaply and fast.
It matters because it separates writing from deciding. This means your big models only do the hard writing and thinking, while Jev handles all the tiny choices. You save tokens, money, and time.
For business owners, the confidence score solves a trust problem. You know when the AI is unsure, so you can review only the uncertain items. This lets you delegate more work to agents safely.
