01What Jev AI Is and How It Works
Jev AI is a new kind of model that does not write sentences. Instead, it picks one answer from a list you give it. You provide a situation, a question, and a set of allowed answers.
Jev chooses one and returns a confidence score between 0 and 1.
This is called a system one model, after the psychologist Daniel Kahneman. System one thinking is fast and automatic, like stopping at a red light. Every other AI model writes its answer word by word, which is slow.
Jev skips all writing and hands you the number directly.
02Main Features of Jev AI
Jev has three question types: choice (pick one from up to 255 options), score (rate on a scale of 2 to 10 levels), and yes or no with a probability. That is the whole surface—simple on purpose.
- It answers in 70 to 500 milliseconds. One independent test measured median 256 milliseconds.
- You can batch multiple questions into one call. Asking 13 questions together is 12.2 times cheaper and 10 times faster than asking them one by one.
- It cannot hallucinate because it only returns answers from your list. It physically cannot type a tool name that does not exist.
- The confidence score is calibrated. When Jev says 0.9, higher accuracy really does follow, because it was trained with a method called calibrated decision reinforcement learning.
03How to Set Up Jev AI
- Go to TypeSafe AI's website and create an account. You need a card on file, but usage stays free until September 25, 2026.
- Pick a coding agent you already use, like Claude, Hermes, OpenClaw, or Codex.
- Give your coding agent the Jev documentation and describe your app in one sentence.
- Define your decision questions. For each one, write a situation, a question, and a list of allowed answers. The description is the most important part—a clear sentence is better than complex code.
- Set a confidence threshold. For low-risk tasks like sorting email, you can act at 0.6. For high-risk tasks like client messages, set it higher and ask for human input when confidence drops below 0.45.
- Test Jev on a small batch of decisions before going to production.
04How to Use Jev AI: Examples from the Episode
Julian shows two real apps built with Jev. The first is a browser agent that controls Chrome by voice. When you say a command, Jev answers about 10 questions at once: what does this person want, where to go, what to click, how far to scroll.
The browser moves while you are still talking.
The second app is a live mirror. You stand in front of a webcam and say where you are going. Jev answers seven questions in under a second: what type of request, which top, which jacket, trousers, shoes, extras, and room.
The app literally changes your outfit on video in real time. The intelligence is just one line of text per clothing item describing when to wear it.
In both apps, Jev also checks its own work. It asks a secondary question like "how well does this outfit match?" If the score is low, it picks again—all in the same half second.
05Results and Comparisons from Tests
TypeSafe AI ran its own benchmark across four business workflows: security alerts, agent finish checks, invoice handling, and customer service. That was 711 total cases. Jev scored 67.8% agreement, taking 0.4 seconds per case and costing 0.04 cents.
GPT 5.6 Tera scored 67.9% but took 10.1 seconds and cost 76 times more. Opus 5 scored higher at 73.1% but took 38 seconds and cost 440 times more. Sonic 5 landed at 67.8%—identical to Jev.
An independent test by the team at Avery ran 777 judgments through Jev in under 0.7 seconds for about a quarter of a cent. In 11 experiments, they made 1,709 judgments for less than one cent total. They sorted startup pitches, graded support replies, flagged risky agent actions, and checked writing for AI tells.
Jev correctly flagged AI‑leaning articles but needed a deeper accuracy check before production.
Near Here, a UK event site, tested Jev on rating listings. Jev got 96% accuracy; Gemini Flash got 86%. Jev was 58 times cheaper per decision.
Julian's own small test: 8 commands, 9 questions per command, median 445 milliseconds, all 8 correct. Claude Haiku took 1.6 seconds, Gemini 3.8 Flash 2.2 seconds, GPT 5.6 Luna 2.5 seconds—and each got only 7 out of 8.
06Tips, Limits and Mistakes to Avoid
Jev cannot write emails, posts, or explain itself. Do not use it for reasoning or hard problems. If you need text generation, pair Jev with a writing model like Claude. Jev handles the decision step; Claude handles the writing.
The speed and cost claims from TypeSafe are their own tests. They used two other AI models to define correct answers, not human judges. So the scores measure agreement, not absolute accuracy.
Treat the benchmark as direction, not a promise.
Jev can give you the wrong answer from inside your list. A confidence threshold helps. For the voice browser, if confidence drops below 0.45, the app puts numbered badges on the page and asks the user which one they meant.
Build a fallback like that into your own workflows.
07Who Jev AI Is For and Why It Matters
Jev is for anyone who makes the same small decisions over and over: sorting leads, routing support tickets, scoring email replies, checking if an agent's output needs a human review. When a decision costs a fraction of a cent and comes back in half a second, you can afford to check every draft, score every lead, grade every output before it goes out.
The founder, Diogo Almeida, came from OpenAI and raised $40 million for this. He says most intelligence should live quietly in the background, not in a chat window. Jev lets you split your work into two models: one that decides and one that writes.
Most agent work was never writing—it was deciding. Now you can pay a fraction of the price and get 25 to 95 times faster decisions.
