On September 15, a San Francisco company called TypeSafe AI emerged from stealth with a 40-million-dollar seed round led by DCVC and a valuation of roughly 200 million dollars. Twenty-four days later, on October 9, the company announced a Series A of 870 million dollars at a valuation of 7.5 billion dollars. The gap between the two numbers is among the steepest on record for an artificial intelligence company, and it took less than a month to cross.
The product behind the jump is called Jev, and it is deliberately unlike the chatbots that have defined the past few years. Jev does not write prose. A developer sends it a block of data and a set of typed questions, and the model returns answers with calibrated probabilities: a choice from a list, a score on a scale, a yes-or-no likelihood. It responds in 70 to 500 milliseconds and prices input at 0.042 dollars per million tokens, with output free.
The price is the sharp edge of the pitch. A frontier language model can cost twenty dollars or more per million input tokens; Jev charges less than five cents, or about a two-hundredth of that, because it emits a decision rather than a paragraph. Output is free on the theory that it is too cheap to meter.
TypeSafe calls Jev a System One model, borrowing the name from the psychologist Daniel Kahneman’s description of fast, intuitive thinking. The company says Jev was trained with a method it calls reinforcement learning for calibrated decisions, aimed at producing reliable probability estimates rather than text that pleases human raters. The distinction is the whole thesis: a decision engine is judged by how accurate its probabilities are, not by how fluent its sentences read.
The company, founded in 2024 by the former OpenAI researcher Diogo Almeida with Erik Gafni and Sasha Sheng, argues that most of the decisions software makes are small and specific, and that a large language model is the wrong tool for them. A model that generates text for a task that needs only a score or a classification is expensive and slow. Returning a typed probability, rather than a sentence, is faster and cheaper and easier for a program to act on.
Andreessen Horowitz led the round, with Sequoia Capital and the existing investor DCVC participating. Martin Casado, an Andreessen Horowitz general partner, is joining the board. The firm’s partners described Jev as a new way to think about how software uses AI, arguing that a model built for cheap, native decisions will be called from code far more often than a model that must be chatted with.
The traction claims are eye-catching for a product three weeks old. The company says about a third of the Fortune 500 are already using Jev. Andreessen Horowitz has said the model generated a trillion tokens in its first three days, a figure that measures invocation volume rather than revenue but that speaks to how quickly developers have plugged the model into existing systems.
The round is a statement about where venture investors believe the next wave of AI value lies. The past three years rewarded companies that could generate text, images and code. Jev’s backers are betting that the next wave rewards models that disappear into the plumbing of software, making thousands of small decisions a day without anyone noticing.
That bet has its skeptics. The valuation is not tied to revenue or profit, and the company has not said whether the 7.5 billion figure counts the new money. An 870-million-dollar check for a product launched a month ago is a wager on velocity, the belief that adoption now will justify the price later. The trillion-token figure, in particular, is a measure of activity, not of income.
TypeSafe has said the capital will fund the development of its System One model family and the enterprise features that paying customers will eventually need. Its documentation describes three question formats: choice, score and a yes-or-no probability it calls noul.
What remains to be seen is whether the decision model can defend its speed and cost advantages as the frontier moves. The large model makers are racing to make their systems cheaper to call, and a general model that can answer a decision question nearly as cheaply would narrow the case for a specialized one. Microsoft introduced a competing decision model on October 9, the same day the round was announced.
For now, the market has rendered a verdict in the only currency venture capital recognizes: a seven-and-a-half-billion-dollar price on a company that was worth two hundred million less than a month ago. The question hanging over the round is whether the fastest repricing in the industry’s recent history will look, in hindsight, like foresight or like a frenzy.
