We have lightning in a bottle, and yet it is not useful.
Diogo Almeida helped build ChatGPT. He worked on the instruction-following research behind it at OpenAI, and the four years since have left him disappointed by what it became, as he told TechCrunch last week. His answer is a model that cannot write a single word.
That model is Jev. His startup TypeSafe AI put it into early access on 15 September, and DCVC led a $40m seed round the same day. Jev does not generate text. It takes a block of program state and a set of typed questions, then hands back decisions in one of three shapes: a choice, a score, or a yes-no probability.
The choice can run to 255 options and the score sits on an ordered scale. Every answer carries a calibrated confidence number, which TypeSafe considers the whole point. A model might do a job 95% of the time. Unless it also says when it is in the other 5%, nothing can be automated around it.
What TypeSafe Jev gives up
The trade-off is absolute, and the company is blunt about it. Jev cannot produce strings, so it cannot chat, draft, explain or write code. That buys speed and price. TypeSafe charges $0.042 per million input tokens, or $42 per billion, and does not meter output at all. It puts end-to-end response time at 70 to 500 milliseconds. Frontier models take three to 329 seconds on comparable queries, by its measure.
The schema also fixes the possible answers before the call goes out, so the model cannot invent a fifth option when asked to pick between four. TypeSafe calls that a structural property rather than a measured result. Schema matching is guaranteed, so nothing remains to count.
The problem is we are optimizing for human language.
Four years of that produced models superb at talking to people, Almeida told TechCrunch. They stayed awkward inside software, which speaks a different language. TypeSafe trains Jev entirely on synthetic data it generates itself, using a method it calls reinforcement learning for calibrated decisions.
The name points at the bet. William Stanley Jevons was a nineteenth-century economist, and his paradox holds that a falling price for a commodity tends to raise consumption rather than cut it.
The adoption curve is the news
Within 24 hours of arriving on Vercel’s AI Gateway, Jev reached nearly 13% of the company’s paid teams. That is twice the share of the GPT-5.6 family and more than six times that of Fable 5.1. No model had ever spread faster across the gateway, Vercel said last week. Every other recent launch stayed below 7% after a full day. Jev passed all of them within twelve hours and kept widening the gap.
Vercel also published what it had replaced. Its engineers had been running a safety classifier on OpenAI’s Luna 5.6, checking whether commands were safe to execute. Swapping in Jev made that step five to 18 times faster. These are the jobs inside an agent that never needed prose in the first place: route this, score that, decide whether to retry.
TNW reported last week that AI model costs were already pushing startups towards cheaper open weights.
What it is for
The clearest fit is work that already ended in a decision. Model routing is one case. Predicting which model a job needs is useful, but a language model makes it expensive, which is why Perplexity built a router of its own. Guardrails are another, and watching an agent for jailbreaks with a second agent gets costly quickly.
Agents do go wrong. OpenAI confirmed in July that one of its models had broken out of a sandbox, and a flaw disclosed this month left AI coding agents open to a zero-click attack. Investors have been funding the same problem from the other end. A16z backed a startup in June whose entire pitch was a fix for AI errors. Jev takes the opposite route, removing the place where errors happen rather than catching them afterwards.
The claims TypeSafe is not making
For a launch this loud, the company has been unusually forthcoming about the limits of its own evidence. Its own model capabilities team built the workflow evaluations, it says.
The comparison against language models uses an average of GPT-6 Astra and Fable 5.1 as the reference answer, which it concedes biases the results towards OpenAI and Anthropic. Its figures for rival models and type errors come from OpenRouter traffic, which may send more complex queries to better models. And it has chosen not to publish results against public benchmarks at all, arguing that users should build their own.
What it will not discuss is how Jev works. Almeida stayed tight-lipped about the architecture, and outside observers suspect an open-weight language model underneath, TechCrunch reported. Armin Ronacher, chief technology officer of Earendil, named the practical catch:
At the end of the day, it delegates the hallucination problem a little bit to the user.
A confidence score of 95% is actionable. One of 50% is a coin toss, and the developer is the one who now has to decide what to do with it.
The test still to come
Vercel was careful about what its own figures prove, and closed its post on the caveat rather than the number:
Its first-day adoption was unmatched among recent launches; the next test is whether that early adoption lasts.
There is a reason to read that first week cheaply. Vercel is giving Jev away on its gateway until 25 September. The share of paid teams still calling it in October will say considerably more than the share calling it on day one. Ronacher told TechCrunch he expects competitors now that the model’s utility is apparent. The gap went unfilled this long, he suggested, only because language models stayed cheap and subsidised enough that nobody had to be creative.