TypeSafe AI's Jev Model Sparks Copycats as Big Tech Joins the Race for LLM Alternatives

It does not chat. It does not write essays or code. And it may be the most talked-about AI release of the fall. Three weeks after startup TypeSafe AI released its Jev model, Silicon Valley is buzzing about alternatives to large language models — and the copycats are already multiplying.

Released on September 15, Jev takes a radically different approach to artificial intelligence. Instead of generating text one word at a time, it takes a piece of context and classifies it into a set of predetermined answers: yes or no, a numerical score, or a selection from a list.

The company describes its training method as "Reinforcement Learning for Calibrated Decisions."

TypeSafe AI is led by Diogo Almeida, a former OpenAI staff member who worked on ChatGPT before leaving in 2024. The company emerged from stealth with Jev's release and has raised $40 million from venture firm DCVC.

Almeida's bet is that the pared-down design makes Jev fast, cheap, and repeatable enough to fix the shortcomings of traditional language models, above all their unpredictability.

In a Monday interview, he said the model is already in use by some 25 percent of Fortune 500 companies and is processing enormous volumes of requests. "We were at a trillion tokens per day about a week ago," he said. "And obviously, growth has been exponential."

The company markets Jev as a better fit for automated software than a general-purpose chatbot. Almeida drew a sharp contrast with today's leading language models, which he said work well with humans in the loop but are "unbelievably useless for automation."

Software, he argued, needs stable layers that developers can build on repeatedly — something chatbots have never provided. The company claims each query costs about 4.2 cents per million tokens and runs roughly 100 times cheaper and faster than leading large language models, with responses in 70 to 500 milliseconds.

The idea has caught on quickly. Within days of the release, independent developers published open-source imitations, including a minimal starter project that gathered more than a thousand stars on GitHub and a lengthy discussion thread on Hacker News.

The rush to replicate the approach suggests a growing appetite among developers for structured, predictable AI outputs in production software.

Large technology companies are now entering the same territory. OpenAI unveiled a limited release of a "Decisions API" at its annual developer event in late September, designed to make fast judgments from a fixed set of options — a product closely resembling Jev.

Amazon Web Services and Cloudflare are also moving into the decision-model space, turning what began as a startup experiment into a multi-party industry race.

Investors are paying attention. The Information reported last week that TypeSafe AI is in talks to raise $1 billion or more, with some prospective investors suggesting a valuation upward of $10 billion. Almeida declined to comment on the reported fundraising.

A partner at DCVC said the startup is already profitable because of its sharply lower computing costs.

Whether Jev-style "decision AI" becomes a durable new category or a feature absorbed into existing platforms, the direction is clear: after years of chasing ever-larger language models, a growing part of the industry is betting that the next wave of software automation will be built on AI that answers instead of talks.