OpenAI says it solved a Millennium problem, and a priority fight has broken out

To put it simply, the Millennium Problems are seven open problems chosen by the Clay Mathematics Institute in 2000, with $1 million awarded for solving any one. The Navier-Stokes problem asks whether the equations that describe the flow of water or air blow up to infinity under certain conditions, or always stay smooth.

The numbers are overwhelming. By OpenAI's account, the work began on September 1 with roughly 10,000 agents working at once, spending 2.7 million messages and about 130 billion output tokens on this problem alone. The analytical proof arrived on September 5, 88 hours after the start, and GPT-6 Astra spent a further 17 hours checking it with the formal verification tool Lean. Fortune put the computing cost at roughly $2 million.

Timing is half of this announcement. OpenAI itself said it began the work on September 1 after rumors circulated that Anthropic's side was about to release a solution. And one day earlier, on September 7, NYU mathematician Tristan Buckmaster issued a statement saying that he, together with Levent Alpöge, who works at Anthropic, had proved finite-time blowup for the Euler equations with help from Claude and Codex, and in that statement he described a phone call with OpenAI.

The two accounts now collide head-on. Buckmaster wrote that in a call on September 6, OpenAI's Sébastien Bubeck proposed dropping Alpöge, an Anthropic employee, from the author list, and that when he refused he was asked why he wanted to ruin his career. In a post on X on September 8, Bubeck countered that he never asked for Alpöge to be removed as an author, and that the conversation was one he initiated to coordinate the two sides' announcement schedules.

Bubeck also denied the data-access allegation. When Buckmaster raised the possibility that OpenAI had seen the work through his Codex account or training data, Bubeck said they never fed the two researchers' prompts or proof into the model or steered the agents in that direction. OpenAI's blog likewise stated that it did not see that work through any channel before the two made it public, and that on the forced Euler problem it recognizes the priority of Alpöge and Buckmaster.

The reaction from mathematicians is cold and heavy. Fields Medalist Terence Tao pointed out that AI companies are using old open problems as marketing evidence, and wrote that if only answers come out and no insight remains, the ecosystem in which the next generation of mathematical techniques grows could collapse. Fortune summed up the controversy in two words, cheating and threats, and which side is right has not been confirmed.

So this announcement has to be read by separating what has been confirmed from what has not. That a proof exists which passed Lean verification, that OpenAI said it will not claim the prize, and that both sides claim not to have seen each other's results: those are the confirmed facts for now. Who was first, and exactly what was said on the call, are points where the two people's records differ and there is no third-party confirmation.

What this means for Korean readers is that the way mathematical research competes has changed. A problem no human had solved in more than a century was pushed through in four days by 10,000 agents, and who announces the result first has become a company's report card. A situation in which a company's announcement schedule, not a research lab's rules, sets the pace of mathematics has actually come to pass.

In the end, this case will be decided by procedure, not by the proof. The equations were verified by Lean, but no one has yet verified anything about priority, the author list, or data access, and the moment the Clay Institute begins its review, that procedure will be entered into the official record for the first time.