OpenAI announced on Sept. 8 that roughly ten thousand AI agents, working together, produced a complete proof of the Navier-Stokes singularity problem13. It is one of six unsolved Millennium Prize problems. Fluid dynamics researchers have chased it for two centuries. The agents worked continuously across eighty-eight hours, from Sept. 1 to Sept. 52. A separate check confirmed the result in the formal proof language Lean. Early credit to Tristan Buckmaster was missing from the celebration. The NYU mathematician had unpublished work with Anthropic researcher Levent Alpöge. It appears to have shaped the direction OpenAI's model took.
Buckmaster put his discomfort on the record: "There is another part of this story, and one that, honestly, I very much wish I did not have to be concerned with"1.
Ten Thousand Agents, One Weekend
Scale defined the approach: engineers set loose linked agent groups, each able to talk inside its own cluster and each pointed at the same unsolved problem2. They aimed at the full three-dimensional Navier-Stokes equations. Mathematicians had spent decades on the easier Euler case. Agents drew on cached web access and code tools. They generated 2.7 million internal messages and roughly 130 billion output tokens before they landed on a singularity proof. GPT-6 Astra then spent seventeen more hours translating the informal argument into Lean, the proof assistant mathematicians now trust to catch the errors human reviewers miss. Force, more than elegance, won. OpenAI's own account stops short of claiming insight into why the equations break down3. It states only that they do.
The Call That Changed Everything
Buckmaster's account begins on Sept. 31. He told an OpenAI mathematician about his and Alpöge's unpublished work. The work stood apart from either employer, he stressed. Three days later Sébastien Bubeck called. OpenAI's internal model had already produced a hundred-page proof of forced Navier-Stokes, he said.
Buckmaster says he told OpenAI of his unpublished work on Sept. 3, while its agents ran Sept. 1 to 5
Dates in OpenAI's Navier-Stokes run and the credit dispute, Sept. 1 to Sept. 8, 2026
Each pin is one dated event on a rail spaced by real time: blue pins are OpenAI's run and announcement, and the magenta pin is Buckmaster's outreach, by his account.
Sources: TechCrunch [1]; Quanta Magazine [2]; Fortune [4]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| Sept. 1, 2026 | OpenAI's agents begin the 88-hour run. Roughly ten thousand agents, working together |
| Sept. 3, 2026 | Buckmaster tells OpenAI of unpublished work. Bubeck says his team first saw the work at its public release |
| Sept. 5, 2026 | The agents finish the run |
| Sept. 8, 2026 | OpenAI announces the proof. A separate check in Lean confirmed the result |
Buckmaster explained why the coincidence troubled him: "The route to the Clay problem through a smooth force, options c and d in Fefferman's statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement"1.
Bubeck offered two paths, by Buckmaster's account1. OpenAI publishes a day behind Buckmaster's team, or Buckmaster alone writes up the result, with Alpöge stripped from the credit line because he works inside a rival lab. Buckmaster refused the second option. Bubeck then asked him, "Why would you ruin your career?". A later line from the same conversation landed harder: "If you don't want me to be nice, then I don't have to be nice".
Bubeck's Defense and Its Tell
Bubeck denied the copying charge directly: "We did not use their prompts or proofs to prompt our models or direct our agents. We, whether it's the researchers or the agents, did not see any of their work until they were released publicly yesterday night"4. Weigh the scope of that denial. It covers prompts, proofs, researchers, and agents, and it ends at a timestamp. What a model absorbed earlier, through other channels, sits outside every clause of it.
He also conceded priority: "I want to be extremely clear that we recognize the priority of Levent Alpöge and Tristan Buckmaster's work"4. Study the timing. That concession arrived after the dispute reached print. Going public did what a private call left undone. A third statement turned to his own position: "A series of false and inflammatory allegations against me are currently circulating on social channels. To clarify, I came into the discussion following academic norms, and I'm disappointed that it has come to this".
What Machines Owe Their Teachers
Buckmaster raised a quieter question beneath the credit fight. He used OpenAI's Codex tool throughout the project1. Then he asked the company what became of those sessions. His account of the answer is exact: "I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer"4.
Read the structure of that reply. One question drew a denial. The other drew silence. A lab built partly on ingesting the open internet faces a new possibility4. Its paying research customers may feed training signal it can barely trace. Every mathematician who runs an unpublished proof through a paid coding assistant now inherits Buckmaster's dilemma, trading convenience for a claim on work that once belonged to them alone.
Tao's Warning and the Ecosystem Question
Terence Tao, widely held to be the era's greatest living mathematician, offered the sharpest outside critique4. His warning was specific: "The indiscriminate strip-mining of open problems for solutions may destroy the ecosystem from which the next generation of mathematical techniques, problems, and practitioners would have developed". Weigh the word ecosystem. Tao describes a field that reproduces itself through unsolved problems. Trophy results take what the next generation would have trained on.
Tao built his own reputation on collaborative, credited mathematics. A scholar known for generous credit now warns about machine speed. It threatens the norms that made his own career legible to peers. Buckmaster framed the sharper stake as training1. A "Deep Blue-Kasparov moment for mathematics" changes how graduate students learn to prove things, as much as who gets credit.
The Compute Behind the Contest
Money sits underneath the human drama. Running ten thousand coordinating agents for eighty-eight hours against frontier compute costs several million dollars by outside estimates2. Only a handful of labs can spend that chasing one unsolved problem for prestige. Buckmaster and Alpöge cracked the narrower Euler case weeks earlier. They used roughly a hundred agents and fifty hours, on a fraction of the budget. Clever aim still beats raw compute on the hardest problems. That holds until someone with more compute learns where to point it.
OpenAI used about 10,000 agents on Navier-Stokes; Buckmaster and Alpöge cracked the Euler case with about 100
AI agents behind each effort, one cube per 100 agents
Each cube stands for 100 agents: the magenta cubes are OpenAI's run on the full Navier-Stokes problem, and the single blue cube is Buckmaster and Alpöge's Euler effort.
Sources: TechCrunch [1]; Quanta Magazine [2]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| OpenAI, full Navier-Stokes | 10,000 | Eighty-eight hours, several million dollars by outside estimates |
| Buckmaster and Alpöge, Euler case | 100 | Fifty hours, weeks earlier, on a fraction of the budget |
Telling the two efforts apart requires exactly the human judgment that made Buckmaster's work valuable, and OpenAI's agents skipped that step because two mathematicians had already supplied it for free2. Every future Millennium Prize attempt now carries a second question alongside its mathematics. Which humans supplied the insight that compute merely scaled, and do those humans get named?
What This Week Actually Settled
Diego Córdoba and Luis Martínez-Zoroa opened the underlying strategy years earlier, and Buckmaster and Alpöge extended it into the Euler result that landed weeks before this one2. Princeton's Charles Fefferman, an authority on the mathematics, called the resolution a moment that left him "thrilled that the problem was solved". Martínez-Zoroa's reaction stayed personal: "I'm very happy for Tristan," he said, then repeated it for emphasis.
Credit chains of this kind run from opening strategy through extension through machine checking, and they used to sort themselves out slowly, in citations and tenure letters written years apart. Compressing that chain into one contested week is the actual innovation Sept. 8 produced. Mathematics now needs new norms fast enough to keep pace with labs that publish on a product timeline.
By the numbers
- Ten thousand: peak concurrent OpenAI agents deployed on the Navier-Stokes proof, working Sept. 1 to Sept. 51.
- Two point seven million: internal messages the agent groups exchanged while producing the proof2.
- Nearly 130 billion: output tokens the effort consumed before formal verification began2.
- Seventeen additional hours: time GPT-6 Astra needed to formalize the result in the Lean proof language2.
- One million dollars: the Millennium Prize purse OpenAI said it declined to claim3.
- Nearly a hundred agents, fifty hours: the smaller effort Buckmaster and Alpöge used to crack the narrower Euler case weeks earlier2.
- Three days: the gap between Buckmaster's private outreach to OpenAI and Bubeck's account of a completed rival proof1.
- Two centuries: roughly how long the Navier-Stokes equations have resisted a complete mathematical account of their behavior3.
What to watch
Bubeck faces continued pressure to publish a complete, verifiable timeline of what his model saw and when1. Mathematicians say that would settle the dispute better than any statement from a press office. Anthropic gains an unplanned recruiting edge4. Alpöge's public vindication makes its research culture the safer place for mathematicians wary of losing credit. Clay Mathematics Institute officials have yet to weigh in3. Does a forced Navier-Stokes result meet the prize committee's original criteria? That ruling shapes how the next AI attempt gets framed.
Sources
- Russell Brandom, "OpenAI Fought Dirty on Career-Making Math Problem, Says NYU Mathematician," TechCrunch, Sept. 8, 2026, https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/
- Quanta Magazine, "AI Has Solved One of Math's $1 Million Millennium Prize Problems," Quanta Magazine, Sept. 8, 2026, https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
- The Next Web, "OpenAI Publishes Its Navier-Stokes Proof, Skips the Millennium Prize Claim," The Next Web, Sept. 8, 2026, https://thenextweb.com/news/openai-navier-stokes-proof-published-millennium-prize
- Fortune, "OpenAI Says It Cracked Navier-Stokes, One of Math's Grand Challenges," Fortune, Sept. 8, 2026, https://fortune.com/2026/09/08/openai-says-it-cracked-navier-stokes-math-grand-challenge-buckmaster-accusation-cheating-intimidation-tao-lament/
- Semafor, "OpenAI Agents Find Proof to $1 Million Millennium Prize Problem," Semafor, Sept. 8, 2026, https://www.semafor.com/article/09/08/2026/openai-agents-find-proof-to-1-million-millennium-prize-problem
- OpenAI, "On the Navier-Stokes Millennium Prize Problem," OpenAI, Sept. 8, 2026, https://openai.com/index/navier-stokes-solution/