Ex-OpenAI Researcher Launches Jev: A Text-Free AI Model That Surprises Developers With Speed and Low Cost

Not a language model, but a decision machine

A new AI model from company TypeSafe AI is generating notable enthusiasm among software developers. The model, called Jev, fundamentally differs from well-known large language models (LLMs) such as ChatGPT. Instead of generating text, Jev produces probabilistic outcomes — what the company calls 'calibrated decisions.' According to TypeSafe, this makes it unsuitable for open-ended conversation, but very well suited to specific, well-defined tasks within software automation.

The model is transformer-based, like most current language models, but is trained exclusively on synthetic data using a technique founder Diogo Almeida calls 'reinforcement learning from calibrated decisions.' Because users define the possible outcomes in advance, the company says Jev cannot hallucinate — a problem LLMs still struggle with. This was reported on September 18, 2026, by TechCrunch journalist Tim Fernholz.

From OpenAI researcher to founder

Almeida helped develop ChatGPT at OpenAI and co-developed reinforcement learning from human feedback (RLHF), a technique underlying virtually all modern language models. Two years ago he founded TypeSafe AI, he says out of frustration that language models, despite their capabilities, are not practical enough for automation. Computers, he told TechCrunch, 'speak a different language' than humans.

That frustration led to a radically different approach: no language, but numbers. TypeSafe describes Jev as a 'System One' model — referring to intuitive, fast thinking rather than step-by-step reasoning. The name Jev references the nineteenth-century economist William Stanley Jevons, known for the Jevons paradox: the idea that falling costs of a product actually lead to increased use. Almeida hopes for a similar effect for artificial intelligence.

Striking pricing model and high demand

Jev's pricing model differs sharply from common LLMs. Output tokens are free; only input tokens are charged, and billing is per billion tokens rather than per million as with most language models. This points to significantly lower costs per use.

Demand for the model proved so high that TypeSafe temporarily had to stop offering API access to new users. The company has not disclosed Jev's exact architecture; outside observers suspect it is built on an existing open-weight language model.

Developers' practical experiences

At infrastructure company Vercel, Jev was deployed as a replacement for an OpenAI model that served as a safety classifier for commands. According to software engineer Pranit Sharma, the switch delivered results five to eighteen times faster, with higher accuracy.

At Bryo AI, CTO Nikhil Mudholkar tested Jev against Google's Gemini for classifying business emails. Gemini proved slightly more accurate, but ten to twenty times more expensive. Mudholkar particularly praised Jev's reliable probability scores as well suited for automating workflows.

A complement to existing AI models

According to Almeida, Jev can also serve as a control mechanism alongside LLMs, for example to monitor the behavior of AI agents and prevent jailbreaks. Armin Ronacher, CTO of Earendil and creator of the open source tool Pi, points to the possibility of 'model routing': determining which AI model is best suited to perform a task — something that would be unfeasible with expensive language models, but becomes realistic given Jev's low cost and speed.

Ronacher expects competitors to follow with similar models now that the usefulness of this type of technology has become clear. He notes that language models have so far been relatively cheap and subsidized, reducing the need for this kind of innovation.

Next steps

Almeida calls the decision to build entirely on proprietary synthetic data one of the best decisions of his career — more important, he says, than the launch of TypeSafe itself or his work on RLHF. He says that half the company functions as a 'lab' focused on statistically grounded synthetic data. TypeSafe wants to develop more versions of the model, in new modalities. Asked whether this makes TypeSafe a 'frontier lab,' Almeida, according to the article, gave no definitive answer.

Concrete figures on pricing, revenue, user numbers, or investors have not been disclosed by the company.