OpenAI used a $22.5 million compute sprint to outrun the mathematicians whose work it depended on

Twenty-five Fields Medal winners just signed an open letter that should embarrass OpenAI. The signatories include Terence Tao, Pierre Deligne, Maryna Viazovska, Peter Scholze, and Yu Deng, who received his medal earlier this year. They are the closest thing mathematics has to a supreme court, and they used their names to say, in public, that AI labs are threatening the intellectual culture of mathematics. They did not have to do that. They did it anyway.

The trigger is OpenAI’s behavior around the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize problems the Clay Mathematics Institute has been dangling a $1 million bounty on since 2000. The fluid equations that describe how air moves over a wing and how blood moves through an artery have been unresolved in their theoretical form for roughly 90 years. In a single week in early September, OpenAI decided it was going to crack one of them.

What OpenAI actually did

On September 1, according to OpenAI’s own blog post, the company heard “rumors that two Millennium Prize problems had been resolved.” Within hours, OpenAI pivoted an internal AI model “significantly more capable than GPT-6 Astra” onto the Navier-Stokes problem. By Saturday, September 5, after about 88 hours of work, the system had produced an analytical proof. Lean formalization and verification took another 17 hours. OpenAI published the result on September 8 (OpenAI blog).

The compute cost is the part that matters. OpenAI says the Navier-Stokes resolution used about 10,000 concurrent AI agents, 2.7 million messages, and 130 billion output tokens. Across all the Millennium Prize problems it tried, the total was 300 billion tokens and 4.9 million messages. At OpenAI’s current listed API rates, the full project cost about $22.5 million of compute (TechCrunch, September 11).

That is roughly the entire research budget of a mid-sized mathematics department, burned through in a long weekend, by a system that did not know the problem existed until it heard a rumor.

The “rumors” OpenAI heard turned out to be about Levent Alpöge, an Anthropic mathematician, and Tristan Buckmaster, a professor at NYU’s Courant Institute. They had been working on a related Millennium problem and announced three proofs on September 8, the same day OpenAI published. In his statement on the proofs, Buckmaster said OpenAI asked him not to credit Alpöge on the work that had seeded the breakthrough, and that when pressed about when OpenAI started and how much human input was involved, “the answers became more evasive.” Eventually, he wrote, “it was agreed that [the first prompt] had been sent in the past few days, after information about our work had reached OpenAI” (Buckmaster statement PDF; TechCrunch, September 8).

OpenAI’s post denies seeing any of Buckmaster’s work before the public release and says the company offered a joint announcement recognizing Buckmaster and Alpöge’s priority. Two accounts of the same week, sharply different. The 25 Fields Medalists sided with Buckmaster.

What the medalists actually said

The declaration at mathandai.org is restrained and specific. The signatories do not argue that AI cannot help with mathematics. They argue that the way OpenAI and its peers are using it is destroying the cultural infrastructure that makes mathematical work possible.

“Often these solutions are announced in a rush,” they wrote, “leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others. As in all creative professions, this raises severe attribution and plagiarism questions.” The bigger problem, they wrote, is what happens next: “Without the willing mathematicians who must take care of their development and integration into the mathematical canon, AI-conceived ideas would never become fully alive and the crucial human transmission chain between mathematicians would be lost.”

The medalists’ argument is about who does the work and who gets the credit, not about whether AI can help mathematics at all. They are saying that the slow, public, attributable process of human mathematical research is itself a thing of value, and that AI labs are trying to skip it without paying for it.

The steelmanned case for OpenAI

The strongest version of OpenAI’s side is that the proof itself is real and useful. A formal, Lean-verified solution to a statement of a Millennium Prize problem is a genuine scientific achievement. Mathematicians have been stuck on the Navier-Stokes question for nearly a century. If a $22.5 million compute sprint can move the needle on a problem that has resisted decades of human effort, that is something worth celebrating, not attacking. The medalists themselves have signed documents calling for AI assistance in mathematics, including the Leiden Declaration of June 2026, which the International Mathematical Union endorsed.

OpenAI also points out that it offered Buckmaster and Alpöge a joint announcement recognizing their priority. If the labs and the math community are both trying to advance the same work, the argument goes, then public conflict between them is just noise.

The problem with that case is that it ignores who holds the leverage. Buckmaster and Alpöge can offer OpenAI a joint announcement or refuse one. OpenAI can offer to spend $22.5 million of compute on whatever problem Buckmaster and Alpöge are working on, the moment they hear a rumor about progress. Those are not symmetric positions. The human researchers cannot race the lab. They can only hope the lab decides to be gracious, which Buckmaster’s statement suggests they were not.

The bigger pattern

A day before Buckmaster published, Jacob Coxon, a researcher who had spent the last three years doing pretraining work at both OpenAI and Anthropic, resigned and posted his reasons publicly. “Neither company is acting responsibly,” he wrote. “They are racing straight to self-improving superintelligence and gambling with our lives” (Coxon post on X).

On September 10, OpenAI’s Dan Roberts announced the company was withdrawing its sponsorship of the Caltech Mathathon “after considering concerns raised by members of the mathematics community” (Roberts post on X). That is the standard playbook: sponsor the field, then quietly pull the sponsorship when the field objects to how you are treating it.

The compute asymmetry is the structural problem. A frontier lab can out-spend any individual mathematician by five or six orders of magnitude, run the problem while the human is still writing up the result, and then offer a “joint announcement” on the lab’s preferred terms. The medalists are right that this threatens the open culture of mathematics, not because the proofs are wrong but because the proofs are now happening in a system that has no place for the humans whose work seeded them.

Mathematics is the canary. The same dynamic is already running in software engineering, in scientific research, in design, in writing, in every creative field where the inputs to AI systems are the work of human beings who are not being paid or credited for what they contributed. The labs get the breakthrough. The humans get the citation, when they are lucky, or the press release that says “in collaboration with” without ever naming them.

Twenty-five Fields Medal winners signing their names to that critique is a public warning from the most authoritative group of mathematicians alive. The labs are eating the field, and the medalists want the public to see it happening. OpenAI is welcome to spend $22.5 million proving more theorems. It is not welcome to spend that money racing the mathematicians whose work it depends on, pressuring them on attribution when caught, and pulling the sponsorship when the field complains. OpenAI calls it collaboration. It looks like conquest with a press release.