Jev’s breakout has turned routine AI decisions into a three-way fight
Jev’s breakout has turned routine AI decisions into a three-way fight
Jev’s rise began with a simple bet: businesses do not always need an AI to write an answer — often they need it to make a quick, bounded call. After TypeSafe AI released the model on Sept. 15, developers used it to sort emails, classify documents and route customer-service requests, attracted by its promise of cheaper and faster decisions than general-purpose language models.
TypeSafe’s case is that this is automation, not conversation. Co-founder and CEO Diogo Almeida, a former OpenAI researcher, said language models have been strong at human language but are poorly matched to automation because “computers speak a different language.” Jev returns probabilities, scores or preset choices rather than generated text — a design meant to let software act automatically when confidence clears a threshold, while passing uncertain cases to people.
The market moved quickly. Within a day on Vercel’s AI Gateway, nearly 13% of paid teams had tried Jev, and on Oct. 6 OpenAI launched its competing Decisions API. OpenAI’s product uses GPT-6 Luna, packaged and tuned to deliver structured, near-real-time decisions; the company says the dedicated interface can be up to 10 times faster than using its general-purpose Responses API. OpenAI’s Nikunj Handa said Jev had “inspir[ed] this whole thing,” and that the project had not been on the roadmap four weeks earlier.
TypeSafe then turned the momentum into an $870 million raise at a $7.5 billion valuation, saying a third of the Fortune 500 was already using Jev. The backing underlines how consequential tiny differences in latency, price and error rate can become when firms automate thousands of daily judgments.
But a third front is pressing the control question. A widely shared post claimed the open-weight pplx-decider v1.1 got 643 of 669 clinical decisions correct, against Jev’s 628, at 42% less per decision and similar speed — while allowing hospitals to run the model internally. The figures are a claim in a social-media post, not an independently reported benchmark, but they sharpen the contest: proprietary convenience versus open deployment in high-stakes settings.
Write a comment