TypeSafe AI launched Jev on September 15, 2026, and calls it a “System One” model, a name borrowed from Daniel Kahneman’s fast, intuitive kind of thinking. Jev doesn’t write. You hand it state, like a support ticket, a log line or a form submission, and a typed question: yes or no, pick one of these, or give it a score. It hands back a probability, usually in under half a second, at $0.042 per million input tokens with output free. In practice that’s a few cents per thousand decisions. The company’s founder, Diogo Almeida, helped build the training methods behind ChatGPT, and TypeSafe says it trained Jev with a new method it calls RLCD, reinforcement learning for calibrated decisions.
The launch post makes big claims and, to its credit, flags many of its own caveats. Its own evaluations show Jev up to about 200 times faster and 444 times cheaper than frontier models. Independent tests land closer to 5 to 25 times, depending on what you compare it with. “Can’t hallucinate” means the answer always fits the format you asked for, not that it’s always right. It’s weak on counting, math and dates unless those are handled in code first. And the headline claim, that its probabilities are calibrated, meaning its 80% answers come true about 80% of the time, has no published paper or evidence behind it yet.
Here’s why it matters for the businesses we build for. Most of what an AI agent does all day isn’t writing. It’s deciding: is this lead real, which queue does this ticket belong in, is this task finished. The usual way to get those answers has been to pay a large language model to write a paragraph, then pull a yes out of it. A model that returns the decision directly, fast and cheap, changes the shape of that work. Decisions become typed, logged and checkable, and the language model is saved for the parts that actually need words.
For clients, the value depends on volume and on trust. At thousands of decisions a day, in routing, deduplication or spam triage, the savings are real. For a small business making a few hundred decisions a day, the bill barely moves, and a wrong answer costs more than a slow one. The feature that would make Jev worth adopting is the calibration claim, and it’s the claim with the least proof. So our read is to run it beside what a client already has, where a wrong answer is cheap and reversible, keep a fallback, and switch only where it wins on their own data.
Access is still narrow. Jev has been on Vercel’s AI Gateway since September 16, where roughly 13% of paid teams tried it in the first day, and it’s also on OpenRouter and Cloudflare. TypeSafe paused direct signups on September 22 because of demand. We’ll test it on our own contact form, scoring leads beside the rules we already run and never routing on it, and publish what it shows, misses included.
Source TypeSafe AI: Introducing System One models and Jev
