Diogo Almeida helped build the methods that taught language models to follow instructions and talk to people, research that ended up behind ChatGPT. On September 15 his startup, TypeSafe, released a model that refuses to talk at all. Jev doesn't write sentences. You send it a situation and a list of questions, and 70 to 500 milliseconds later it sends back typed answers with probabilities attached.
Almeida's starting question is a good one: if models have been superhuman at chat for years, where's all the automation? His answer is that chat is the wrong interface for software. Software doesn't want a paragraph. It wants a value it can branch on.
TypeSafe calls Jev a System One model, after Daniel Kahneman's "fast thinking," which is a slightly grand name for something refreshingly plain. There are three question types: yes or no, a pick from a fixed list, or a score. Every answer comes with a probability, and TypeSafe adds a confidence measure on top. Because the possible answers are defined in advance, the model can't invent a fourth department or a tool that doesn't exist. That's where the "can't hallucinate" claim comes from, and it needs fine print. Jev can still pick the wrong option from the list. A routing model that's 90 percent sure a billing ticket is a legal threat will be wrong about one time in ten. TypeSafe says as much: higher confidence is supposed to mean higher accuracy, and you're meant to check that on your own data.
The price is the other surprise. Input costs $0.042 per million tokens, or $42 per billion, and output is free. TypeSafe also says, to its credit, that it can't prove the price isn't subsidized. That sentence deserves to travel with every glowing write-up.
Most of our readers will never call an API, so why care? Because the tools you already pay for are full of small hidden judgments. Support apps, invoice scanners, meeting notetakers and jailbreak filters all keep asking whether something is urgent, a refund or spam. Today those questions often go to a chat model whose answer then has to be parsed back into something a program can use. Jev skips that step. TechCrunch reported that Pranit Sharma, a software engineer at Vercel, swapped a GPT-5.6 Luna safety classifier for Jev and got results five to 18 times faster, with better accuracy. We haven't reproduced that. If it holds up elsewhere, you'll notice it as an app that used to hang for three seconds and suddenly doesn't.
Treat the homepage numbers with more care. "193.6 times faster" and "444.6 times cheaper" come from TypeSafe's own workflow evaluations, built by its own team, which use the average of GPT-6 Astra and Fable 5.1 as the reference answer. TypeSafe itself says those gains are probably at the high end. The claims that are easy to check, like speed, price and whether answers match the schema, are the ones it invites you to test.
The Doom demo is fun and slightly beside the point. Playing a game from structured state at ten queries a second shows off latency, and says nothing about whether a hospital intake form or a school office inbox should be automated at that speed. Jev also won't write your email, fix your spreadsheet or replace ChatGPT. If your work is drafting, comparing or explaining, you still need a model that talks.
Our take: The name is a nice touch. William Stanley Jevons was the economist who noticed that cheaper coal meant more coal got burned, and TypeSafe is betting cheap decisions will go the same way. Whether that's good news depends on how well calibrated those decisions are. So if you get in, the first test to run is calibration: collect a few hundred of Jev's answers where you already know the right one, group them by stated confidence, and check that the 90 percent answers are right about nine times in ten.
Jev is in early access, with a waitlist at typesafe.ai and a route through Vercel's AI Gateway. We haven't been let into the API for this piece.
