Jev: A New AI Model That Refuses to Write Text Has Developers Hooked
Set Trending Topics as a preferred source on Google.
A San Francisco startup released an AI model last week that has no interest in talking to humans. It is called Jev, it comes from TypeSafe AI, and within days it became the most discussed model in the developer community. Demand ran so high that the company’s API briefly stopped serving requests.
What Jev Actually Does
Large language models produce text, one token at a time, each conditioned on the last. Jev works on a different principle. Developers define an output schema in advance, meaning the structure of every possible answer, and then feed in unstructured state: a support ticket, a log entry, a game position. What comes back are typed values with attached probabilities that software can consume directly. TypeSafe calls these “calibrated decisions” and describes the model in its launch post as a “frontier-intelligence function call.”
The name of the model class, System One Models, borrows from Daniel Kahneman’s “Thinking, Fast and Slow,” where System 1 stands for fast, intuitive judgment and System 2 for slow, deliberate reasoning. The model itself is named after the economist William Stanley Jevons, whose paradox describes how falling costs for a resource tend to drive up total consumption.
A representative use case from TypeSafe’s own documentation is customer support routing. Jev decides whether an incoming request goes to ordinary application code, to a language model, or to a human agent. The company also points to tool selection for AI agents, guardrails that screen the inputs and outputs of other models, and classification across very large datasets.
Who Is Behind It
TypeSafe AI was founded in 2024 by Diogo Almeida, previously a researcher at OpenAI and a co-author of the InstructGPT paper that laid groundwork for ChatGPT. Almeida worked on reinforcement learning from human feedback (RLHF), the training method behind today’s chatbot boom. His motivation for leaving is blunt, as he told TechCrunch: “We have lightning in a bottle, and yet it is not useful.” Models have been excellent at human language for four years, he argues, and that skill does little for automation because computers speak a different language.
His co-founders are Erik Gafni and Sasha Sheng. After roughly two years in stealth, the company surfaced alongside the model launch with a seed round of 40 million dollars (around 34 million euros) led by DCVC. Forbes reported a valuation of about 200 million dollars (around 171 million euros), as SiliconANGLE summarized.
Why Developers Are Excited
Three things drive the enthusiasm. The first is speed. TypeSafe reports end-to-end response times of 70 to 500 milliseconds, achieved by generating every output in parallel within a single query. Frontier models, by the company’s own comparison, take between 3 and 329 seconds.
The second is price. Input costs 0.042 dollars per million tokens (around 0.036 euros), and output tokens are free. A demo bot that plays the video game Doom at ten queries per second therefore runs at roughly 7 dollars (around 6 euros) per hour, a stunt The Register also covered.
The third is structural reliability. Because the output schema is fixed up front, type errors are impossible, and hallucinated tool calls or invented categories disappear as a failure mode. Early field reports sound correspondingly upbeat. An engineer at Vercel swapped an OpenAI model out of a safety classifier and reported results five to 18 times faster with better accuracy. Another developer benchmarked Jev against Gemini on business email classification: Gemini came out slightly more accurate while costing 10 to 20 times more.
The confidence scores draw particular praise. Armin Ronacher, CTO of Earendil, described the effect to TechCrunch in dry terms: the model delegates part of the hallucination problem to the user, who now has to decide what probability threshold is worth acting on. Ronacher also sees model routing as a natural fit and expects competitors to appear quickly.
What Remains Open
The most striking numbers so far come from the vendor. TypeSafe published its own workflow evaluations and documented the methodology, but it uses the averaged outputs of GPT-6 Astra and Claude Fable 5.1 as the reference for a correct answer. The company itself acknowledges that this biases results toward the large US labs and that the workflows were built by its own teams. Independent verification is still missing, and the figures will vary considerably by workload, network location, and comparison method.
There are substantive limits too. Schema compliance guarantees the shape of an answer while saying nothing about whether that answer is right. How closely the confidence scores track real accuracy is something teams have to measure on their own data. Jev supports choice sets up to a cardinality of 255, so larger decision spaces require a two-stage scoring approach. Images and other modalities are out of scope for now.
On top of that sits a transparency question raised by heise online: Jev returns no verbal justification for its decisions. The deeper such models are buried inside software, the more pressing it becomes to ask how their behavior can be audited and controlled.
The architecture remains undisclosed as well. Almeida is tight-lipped, and outside observers suspect an open-weight language model at the base. According to the company, Jev was trained exclusively on synthetic data. Whether the aggressive pricing holds up without cross-subsidy will take time to establish, and TypeSafe concedes in its own post that it cannot prove the prices are unsubsidized.
Jev is currently in early access through a hosted API. The company is working through its waitlist and has announced further models in additional modalities.
