OpenAI Math breakthrough draws questions about credit and trust

A claim that artificial intelligence has cracked one of mathematics’ most famous problems should have been a scientific triumph. Instead, it has become an argument about credit, whether private work entered into AI tools can influence a laboratory’s research, and how competing companies should behave when the same discovery is within reach.

At the center is the Navier-Stokes existence and smoothness problem. The equations describe fluid motion, from air over an aircraft wing to blood in an artery. The discussion going around offers a kitchen-table version: they are the “ultimate rulebook” for predicting how milk swirls through coffee, or how any liquid or gas flows. Engineers use numerical versions daily. The unresolved question is whether a smooth three-dimensional flow can become singular, with velocity growing without bound in finite time.

The Clay Mathematics Institute chose Navier-Stokes as one of seven Millennium Prize Problems in 2000 and offered $1 million for a valid solution. Only the Poincaré conjecture has been officially resolved. OpenAI says its proof shows smooth, forced Navier-Stokes flows can blow up, satisfying options C and D in the problem statement. That claim is historic but unsettled. Clay still lists the problem as unsolved, and its rules require publication, two years of scrutiny and broad acceptance before considering a prize.

The research that came first

The dispute began with work by New York University mathematician Tristan Buckmaster and Levent Alpöge, a mathematician employed by Anthropic. Their collaboration was personal, not an Anthropic project. They used public language models, including Claude and OpenAI’s Codex, to extend a research program developed by Diego Córdoba and Luis Martínez-Zoroa. That program constructed fluid blowups with rough external forces; Buckmaster and Alpöge sought the harder step of keeping the forcing smooth.

Their results were important but different from OpenAI’s claim. They reported finite-time blowup with smooth forcing for three related systems: incompressible porous media, the Boussinesq equations and three-dimensional incompressible Euler. Buckmaster said the Euler and Boussinesq breakthroughs arrived on Aug. 15, 2026, and were verified in Lean on Aug. 22. Lean checks whether each formal step follows from encoded assumptions. It can expose logical gaps, although a certificate cannot guarantee the theorem was encoded as intended or replace expert review.

The pair released quickly after hearing that word of their progress had reached OpenAI. Buckmaster apologized for the papers’ presentation, calling the Euler write-up “AI slop.” They would normally have spent weeks turning model output into readable mathematics. The rush matters because credit depends on a dated record, while clarity lets other mathematicians test a difficult proof.

The timeline and competing accounts

OpenAI says it began work on the Millennium problems on Sept. 1 after hearing online rumors of progress. The company says roughly 10,000 concurrent agents worked on Navier-Stokes, exchanging 2.7 million messages and producing about 130 billion output tokens. The system reached its result in 88 hours; GPT-6 Astra spent another 17 hours checking it in Lean. The figures show a new imbalance: a small academic collaboration can suddenly face industrial-scale computation.

Buckmaster contacted OpenAI on Sept. 3 to clarify that the work was personal and concerned related equations. He says OpenAI representatives told him on Sept. 6 that their model had produced a roughly 100-page Navier-Stokes proof. The timing and choice of smooth forcing troubled him because his collaboration had pursued that narrow route. He asked whether the model had been trained on, or could access, Codex sessions containing their drafts.

The distinction is crucial: Buckmaster raised a question, not proof of data theft. He says he does not know whether the data was used and did not see OpenAI’s internal process. OpenAI says neither researchers nor agents accessed the pair’s specific work before release. It cannot completely exclude the possibility that de-identified usage data improved its models. The company says its proof differs substantially from the academics’ results.

The conversation then moved to credit. According to Buckmaster, OpenAI proposed coordinated announcements or a paper in which he presented the Navier-Stokes result while Alpöge was excluded because he worked at Anthropic. When Buckmaster threatened to disclose the timeline, he says an OpenAI researcher replied, “Why would you ruin your career?” and, “If you don’t want me to be nice, then I don’t have to be nice either.” Sébastien Bubeck, identified in Buckmaster’s statement, disputes parts of the account, including that he sought to remove Alpöge, and calls his wording poorly chosen.

What researchers and commentators are saying

Reaction available for now focuses less on the equations than on the rules around them. Thomas Wolf asked, “WTF is this way to handle mathematicians’ work and scientific communication?” Susan Zhang framed the fear: “Can these labs steal your work and scoop you when the stakes are high enough?” That remains unanswered, but it could deter researchers from placing unpublished ideas or valuable drafts into hosted systems. That question reaches beyond mathematics and academic prestige.

Schölto Douglas said it was “extremely sad” that the episode did not become an example of laboratories coordinating, because future stakes would be higher. Commentators speculate that competition before possible initial public offerings drove the rush.

The mathematics and ethics will now be tested separately. Specialists must examine whether OpenAI’s analytical proof and Lean formalization satisfy the Clay formulation. Meanwhile, AI providers and universities face questions that cannot wait two years: what happens to confidential prompts, how model-improvement data is separated from active research, and how credit is assigned when human insight, public scholarship and millions of machine-generated steps converge.

Two conclusions can coexist. AI can accelerate frontier mathematics at a scale that recently sounded implausible. The provenance of that acceleration can be hard to audit when the strongest models are private. The result may stand as a landmark. The dispute is already a warning that future breakthroughs need transparent timelines, enforceable data boundaries and credit rules strong enough to preserve research trust.