TypeSafe Releases Decision Model Jev

Diogo Almeida is one of the researchers behind ChatGPT. At OpenAI he built the methods that made language models follow what people ask, work that became the research behind ChatGPT, and he says he was disappointed by the result. According to press reports he put it this way: "We have lightning in a bottle, and yet it is not useful." TypeSafe AI, the company he left to start two years ago, released Jev on September 15 out of that disappointment. It is a model that produces no text and returns only decisions.

Four days after launch, Vercel announced that Jev had become the fastest-adopted model in the history of its AI Gateway. Within 24 hours more than twice as many paid teams used it as for any previous model; it passed every comparison model within 12 hours; and by hour 24 roughly 13% of paid teams were on it. Vercel wrote that this was double the GPT-5.6 family and more than six times Fable 5.1's share. Other recent launches stayed below 7% even after a full day.

What Jev does is choose. An application sends state and a set of questions, and the model evaluates them all in parallel, returning choices, scores or true-and-false answers with probabilities attached. It accepts up to 255 options, and answers cannot fall outside the format defined in advance. An agent picking one of four tools cannot invent a fifth.

Price shows the character of this model better than anything else. Input costs $0.042 per million tokens and output tokens are free. The company's announcement put existing language models at $0.20 to $10 per million input tokens, with output around five times the input price. Response time runs from 70 to 500 milliseconds, against the 3 to 329 seconds end-to-end that the announcement listed for frontier models.

Almeida opened the announcement with a question: "Models have been superhuman at chat for years, so where is all the automation?" His answer is that the optimization target was misaligned. A model trained to be good at human language is strong when talking with people and becomes a bottleneck inside code. TypeSafe calls its new training method Reinforcement Learning for Calibrated Decisions, and says the training data is entirely synthetic data it made itself.

Even the name comes from economics. Jev is taken from the 19th-century economist William Stanley Jevons, and the company applies his paradox, that cheaper coal increased coal consumption, directly to intelligence. The class name, System One Models, comes from Daniel Kahneman's split between fast intuitive thinking and slow deliberate reasoning. Every order-of-magnitude drop in the price of intelligence unlocks many times more uses, the company argues.

Numbers from the field are testing that premise. According to press reports, Vercel software engineer Pranit Sharma said that swapping the classifier checking command safety from an existing language model to Jev returned results five to 18 times faster and with greater accuracy. Bryo AI chief technology officer Nikhil Mudholkar said that for classifying business email Gemini was slightly more accurate but 10 to 20 times more expensive, adding that Jev "is the only one that hands back a real probability, which makes it ideal for automating workflows." One developer cleared 9,081 product-matching pairs that had been parked for human review in 13 minutes for 32 cents.

Returning probabilities moves where judgment sits. Armin Ronacher, chief technology officer of Earendil, which builds an open-source model harness, said that "at the end of the day, it delegates the hallucination problem a little bit to the user," explaining that an answer that comes back at 50% is a coin toss to be discarded while 95% is something you can act on. Because confidence is calibrated to track accuracy, code handles everything above the threshold and everything below goes to a person, which is how the company designed it to work.

The claim of being 193.6 times faster and 444.6 times cheaper comes from the company's own workflow evaluations. The announcement states that the reference answers were the average of GPT-6 Astra and Fable 5.1, which tilts results toward those two companies' models, and that the evaluation tasks were built by people on its own model capabilities team, so bias may exist. It also notes that speed was measured from laptops against its service on the US West Coast.

The money has arrived too. According to press reports a $40 million seed round led by DCVC backed Almeida and co-founders Erik Gafni and Sasha Sheng, and DCVC general partner James Hardiman described the bet as turning capable models into components developers can build into products. Within three days Vercel, Cloudflare, LangChain and Langfuse had added it, and demand was heavy enough that the company briefly could not serve users through its API.

The place where agents call language models already had a price on it. METAL reported Yoshua Bengio's analysis of why AI agents lie and collude, and much of that risk comes from the fact that a model can produce anything as text. The TypeSafe announcement METAL checked says stepping outside the format is mathematically impossible, and the price of that is giving up text generation entirely.

The question now is not how fast Jev is. Vercel wrote that first-day adoption outpaced any recent launch, then added that the next test is whether that adoption holds. How far software will be allowed to take over judgments whose price has fallen close to zero, and who is accountable when a judgment goes wrong, are not yet written next to the probability numbers.