How Jev was built: a model that only decides, never writes
The story of Diogo Almeida and TypeSafe releasing Jev, a model that only picks from preset answers and returns a probability, far faster and cheaper than chat models.
AN OPENAI RESEARCHER THOUGHT HIS IDEA WOULD TAKE A WEEK. IT TOOK TWO YEARS. THREE WEEKS AFTER LAUNCH, OPENAI SHIPPED ITS OWN VERSION AND CREDITED HIM. His name is Diogo Almeida. He co-authored InstructGPT, the paper behind the way ChatGPT was trained. He was on the team that pioneered post-training at OpenAI. At some point he had an idea that did not fit there. According to Fortune, Sam Altman told him to stop working on RLHF and go build it. He thought making it reliable would take about a week. It took two years in stealth. On September 15, 2026, his company TypeSafe AI released Jev. Jev is not a chatbot. It does not write text at all. You tell it the possible answers in advance, and it returns a decision with a probability attached. Yes or no. Which of these five. How confident. Because it can only pick from the answers you gave it, it cannot make one up. He calls it a System One model, after Daniel Kahneman's fast, intuitive thinking. The name comes from William Stanley Jevons, the economist who noticed that cheaper coal made the world burn more of it, not less. The numbers are the point. A response in 70 to 500 milliseconds, where frontier chat models take 3 to 329 seconds on the same kind of task. Input at $42 per billion tokens. Output free. The launch broke their own servers. Demand was high enough that TypeSafe briefly lost the ability to serve the API. On Vercel, nearly 13 percent of paid teams tried Jev within a day, more than twice as many as any model launch before it. DCVC led a $40 million seed round. On October 6, OpenAI released its Decisions API. OpenAI's Nikunj Handa said it was not even on the roadmap four weeks earlier, and credited Jev with "inspiring this whole thing." OpenAI charges $0.10 per million input tokens. Jev charges $0.042. On October 14, TypeSafe shipped version 1.14 and opened a "danger zone": images, bounding boxes, longer context. Opt-in only, labeled not ready for production. Their benchmark post says the opposite of what every lab says: no leaderboard table, and publish the unflattering results too. He sat down with Latent Space for two hours and twenty minutes to explain it. One line from that interview sums up the bet: intelligence will be more like a database than a coworker. The story is not really about a faster model. It is about a refusal. Everyone was racing to build an AI that talks like a person. He built one that answers like a function. The lesson: in software, the most valuable AI may be the one you never chat with. Refusal to generate text when a decision will do. Refusal to chase leaderboards. Refusal to sell confidence you cannot measure. The interview is free. Most people building agents on $10-per-million-token models will never watch it.