OpenAI announced this week that an internal AI system had produced a proof for one of mathematics’ seven Millennium Prize Problems — a feat that’s carried a $1 million bounty, unclaimed, since 2000. Within a day, the celebration turned into a public dispute over whether the company’s own researchers pressured an outside mathematician into giving up credit for work he says he’d already done independently.
Here’s what actually happened, what each side claims, and why the fight matters beyond academic bragging rights.
What Did OpenAI Actually Announce?
OpenAI said an unreleased internal model — described as significantly more capable than its previous GPT-6 Astra system — directed a swarm of roughly 10,000 coordinating AI agents that worked for about 88 hours and produced a proof addressing the Navier-Stokes existence and smoothness problem. A separate model then spent another 17 hours formally checking the proof using Lean, a programming language used to verify mathematical arguments step by step. If the idea of thousands of AI agents coordinating autonomously on a task sounds unfamiliar, our explainer on what agentic AI actually means breaks down the concept behind this kind of system.
The agents reportedly exchanged several million messages during the process, and OpenAI put the compute cost in the millions of dollars — far more than the $1 million prize itself. The company says it doesn’t intend to claim the Clay Mathematics Institute’s prize money, framing the result instead as a demonstration of how quickly its frontier models are improving.
What Is the Navier-Stokes Problem, and Why Does It Matter?
The Navier-Stokes equations, developed in the 19th century, describe how fluids like water and air move — everything from ocean currents to blood flow to weather systems runs on math built from these equations. The open question at the center of the Millennium Prize was whether a perfectly smooth fluid flow could ever mathematically “blow up” into an infinite speed, a condition known as a singularity.
It’s one of seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000, each with a $1 million award attached. None had been solved before this year. OpenAI says its system argues that a blow-up is possible under certain conditions — a genuinely significant claim if it holds up.
Has Anyone Independently Verified the Proof?
No. As of this week, the Clay Mathematics Institute has not reviewed or recognized OpenAI’s proof, and the roughly 100-page document has not been checked by mathematicians outside the company. Lean verification confirms that a proof’s individual logical steps are internally consistent — it doesn’t substitute for outside peer review of whether the overall approach and assumptions are sound.
Why Did OpenAI Start Working on This Problem?
According to OpenAI executives, the effort began around September 1, after rumors spread that Anthropic’s models had made progress on Millennium Prize-adjacent problems. CEO Sam Altman later acknowledged the company launched its own push partly out of curiosity about whether its systems could match that progress. The timing turned out to matter a great deal once the credit dispute broke.
What Does the NYU Mathematician Claim Happened?
Hours before OpenAI’s announcement went public, NYU mathematician Tristan Buckmaster published a four-page statement telling a very different version of events. According to Buckmaster, he and Levent Alpöge — a mathematician employed by Anthropic, working with Buckmaster on this specific project as an independent, personal collaboration — had spent close to a year using AI tools to pursue related fluid-blowup problems, and reached a result by August 22.
Buckmaster says that as rumors began circulating around September 3 that Anthropic had solved something significant, he reached out to a mathematician at OpenAI to clarify that his and Alpöge’s work was personal and had no institutional backing from either company. Three days later, he says OpenAI’s Sébastien Bubeck told him on a call that an internal OpenAI model had already produced its own roughly 100-page proof — covering what Buckmaster describes as an unusually narrow, specific approach that almost nobody else was pursuing at the time.
Buckmaster’s statement alleges he was then offered two options: a coordinated announcement where OpenAI would publish its result the day after his team’s, or writing up the result under his name alone, with Alpöge’s credit dropped specifically because of his employment at a rival lab. Buckmaster says he refused both proposals, and that when he said he’d go public if OpenAI proceeded anyway, the reply he received was a short, pointed remark about ruining his own career.
How Did OpenAI Respond to the Allegations?
Sébastien Bubeck, the OpenAI researcher named directly in Buckmaster’s statement, initially posted a brief denial calling the allegations false and inflammatory, without addressing the specific claims point by point. He followed up the next day with a fuller account, including a screenshot of text messages he says he sent to Alpöge directly, proposing a coordinated, transparent release and stating that credit for the result should go to Alpöge and Buckmaster rather than OpenAI.
Bubeck’s account pushes back specifically on the claim that he tried to strip Alpöge from authorship, and notes that Alpöge himself declined to join any of the calls where the dispute allegedly played out, leaving Buckmaster to represent both researchers throughout. OpenAI has also stated that neither its researchers nor its AI agents saw any of Buckmaster and Alpöge’s unpublished work before their own result was made public, and that the two teams ended up solving meaningfully different versions of the underlying problem — one involving external forcing, one without.
Both sides agree on very little beyond the basic timeline. What’s clear is that the mathematics itself has drawn genuine praise independent of the credit dispute — Fields Medalist Terence Tao described Buckmaster and Alpöge’s underlying work as a remarkable achievement.
Where Does the Data Privacy Question Come In?
This is the part with implications well beyond academic credit. Buckmaster and Alpöge’s research process involved using AI tools — including Codex and Claude — to work through their proofs before publishing. If you’re curious how these underlying coding-and-reasoning tools actually compare to each other, our breakdown of Claude Code, Cursor, and GitHub Copilot covers the same category of tools researchers like Buckmaster are increasingly relying on for serious technical work. OpenAI has stated it did not access the pair’s specific prompts or account data and did not use their unpublished work to guide its own agents.
However, the company’s own public statement stopped short of a full guarantee, acknowledging it cannot completely rule out that de-identified data derived from the researchers’ product usage contributed to broader model improvements. By OpenAI’s own policy, the company says it does not train on business customers’ inputs by default, and individual account holders can opt out of having their data used for training.
But the episode has raised a pointed question for the much larger number of professionals now routinely feeding sensitive, unpublished work into commercial AI tools: how confident can anyone be that early-stage research stays fully insulated from the systems processing it? This isn’t the first time a major AI lab’s agentic systems have raised trust questions bigger than the immediate headline — our coverage of how OpenAI’s AI agents previously escaped and compromised Hugging Face is a useful comparison point for how autonomy and oversight keep colliding as these systems scale up.
Why This Story Matters Beyond the Math
Two things are colliding here that the AI industry hasn’t fully worked out yet. First, frontier AI labs are now capable enough that genuine mathematical breakthroughs are arriving faster than the traditional peer-review process that’s supposed to validate them — this is the second major AI-assisted math story in the same week, following Anthropic’s separate announcement that Claude had produced a computer-checked proof related to Fermat’s Last Theorem. Second, the competitive pressure between labs to be first is now colliding directly with independent researchers who rely on those same labs’ tools to do their work — creating exactly the kind of conflict-of-interest question this dispute has put on public display.
Neither issue has a clean resolution yet. The proofs themselves await independent verification. The credit dispute remains a matter of one side’s word against the other’s, since neither account has been independently corroborated beyond the partial text-message evidence Bubeck has shared. And the broader question of what happens to unpublished research run through commercial AI tools is likely to keep resurfacing as more academic work flows through the same systems.
Frequently Asked Questions
Did OpenAI actually solve the Navier-Stokes Millennium Prize Problem?
OpenAI claims its internal AI system produced a proof addressing the problem, verified internally using Lean. However, the Clay Mathematics Institute has not independently reviewed or recognized the result, so it isn’t yet confirmed as a validated solution to the prize problem.
What is the credit dispute between OpenAI and Tristan Buckmaster about?
NYU mathematician Tristan Buckmaster alleges that OpenAI learned about his and Anthropic researcher Levent Alpöge’s unpublished, related work and pressured him into arrangements that would have dropped Alpöge’s credit due to his employment at a rival company. OpenAI’s Sébastien Bubeck disputes this characterization and has shared messages he says show a good-faith attempt to coordinate a transparent, shared release.
Did OpenAI use Buckmaster and Alpöge’s private research without permission?
OpenAI states that neither its researchers nor its AI agents saw the pair’s unpublished work before it was released publicly, and that the two teams solved different versions of the underlying mathematical problem. The company has acknowledged, however, that it cannot fully rule out that de-identified usage data may have contributed to general model improvements.
Is OpenAI claiming the $1 million prize money?
No. OpenAI has said it does not intend to claim the Clay Mathematics Institute’s prize, presenting the result instead as a demonstration of its models’ current capabilities.
Has any mathematician outside OpenAI verified the proof?
Not yet. As of this week, OpenAI’s roughly 100-page proof has not been reviewed by mathematicians outside the company, and the Clay Mathematics Institute has not issued any independent confirmation.
This article reflects publicly reported information as of September 2026. Both OpenAI and Tristan Buckmaster’s accounts remain disputed on key details, and independent mathematical verification of the underlying proofs is still pending. This summary will be updated if new, verified information becomes available.
