Diogo Almeida felt let down by ChatGPT. As an OpenAI researcher, he had contributed to building the chatbot and to developing reinforcement learning from human feedback (RLHF)—the training method largely credited with ushering in today's AI era. Yet he found himself dissatisfied. "I've been battling that problem since then. It took me a while to come to the conclusion: the problem is we are optimizing for human language … We have been super good at human language for four years, but it's not useful for automation because computers speak a different language."

Two years ago, Almeida departed OpenAI to found TypeSafe AI, a startup aimed at solving this issue. This week, the company unveiled a new transformer-based model called Jev—one that is decidedly not a large language model (LLM). Rather than generating text, Jev produces probabilities, which the company describes as "calibrated decisions."

By sidestepping language entirely, the model achieves several advantages: it becomes remarkably cheap and fast, and since users predefine the outputs, hallucination is impossible. Output tokens cost nothing, while input tokens are billed by the billion rather than the million.

Developers have shown considerable enthusiasm for the product—demand was so intense that the company temporarily lost its ability to serve users through its API. Jev appears particularly well-suited for software automation, and developers are viewing it as a more affordable and reliable method for embedding intelligence into their code.

Pranit Sharma, a software engineer at Vercel, a company specializing in agentic infrastructure, noted that his firm had previously used OpenAI's ChatGPT Luna 5.6 to operate a classifier for reviewing commands for safety. When Vercel swapped OpenAI's Luna for Jev, the results came back 5 to 18 times faster and with improved accuracy. Another developer, Nikhil Mudholkar, CTO of Bryo AI, benchmarked Jev against Gemini for categorizing business emails. In his evaluation, Gemini was marginally more accurate but cost 10 to 20 times more. What intrigued Mudholkar more were Jev's confidence scores—"it is the only one that hands back a real probability which makes it ideal for automating workflows."

Beyond substituting for LLMs in specific scenarios, the new model can also complement them, serving as an intelligent monitor for misbehavior. Deploying agents to oversee other agents can become costly quickly, but Almeida contends that using Jev for this purpose is practical. He envisions users employing Jev to trace LLM agent activity and guard against jailbreaks.

"At the end of the day, it delegates the hallucination problem a little bit to the user," explained Armin Ronacher, CTO of Earendil, which develops the open-source model harness Pi. "The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it."

Model routing represents another potential application for Jev, according to Ronacher. Determining whether a particular workload needs a specific model would be valuable, but relying on an LLM for that task would prove expensive. Jev's low cost and speed enable that kind of real-time sorting.

That aligns with Almeida's vision. The model takes its name from William Stanley Jevons, the 19th-century economist whose namesake paradox illustrates how declining costs of a commodity can drive increased usage. In this instance, the decreasing cost of intelligence should result in its broad deployment.

"We think that there's just going to be smart software all over the place in a way that's emergent and distributed … much more like the early internet than you know like the the mega apps that people are trying to build right now," Almeida said.

Almeida remains secretive about the model's architecture, which outside observers believe is constructed atop an open-weight LLM. The company characterizes Jev as a "System One model," emphasizing intuition over reasoning and focused on the appropriate task. Almeida states that Jev is trained solely on synthetic data using a method he terms "reinforcement learning from calibrated decisions." "Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data, and that is now my life joy."

Currently, Jev occupies a unique position as this type of model, but Ronacher anticipates competitors will emerge once its usefulness becomes clear. "We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don't have to be creative yet," he said.

TypeSafe intends to develop additional versions of the model in new modalities. When asked whether TypeSafe qualifies as a frontier lab, Almeida responded, "the main product of Frontier Labs is fear or hype. I would like our main product to be intelligence…[but we are] not a lab in the sense of, you know, like bet on infinite wealth, or a religion, or building God in a data center, or whatever is the the thing of today."