On 8 September 2026, OpenAI announced that its AI model Astra produced a proof of the Navier-Stokes Millennium Prize problem in 88 hours, using a swarm of roughly 10,000 autonomous agents. NYU mathematician Tristan Buckmaster immediately accused the company of copying his team’s earlier work on the related Euler problem after learning of their progress.
Buckmaster says the episode reveals a deeper fracture: AI compute has made racing to publish mathematical breakthroughs pointless. He is calling for ground rules between academics and AI labs, warning that without them, the traditional model of open mathematical research is broken.
“I think it’s pointless. Like, I think the game is up.” Tristan Buckmaster, a British-Australian mathematician who trained in Germany before landing at New York University, is not talking about a lost chess match. He is describing the state of mathematical research after OpenAI claimed to have solved a Millennium Prize problem in 88 hours using a swarm of AI agents.
Buckmaster and his collaborator Levent Alpöge, now at Anthropic, had spent months working on the Euler equations, a stepping stone to Navier-Stokes. They reached a solution in August 2026, using AI tools from both Anthropic and ChatGPT. Days later, according to Buckmaster, OpenAI scientists approached him proposing a joint announcement on Navier-Stokes, claiming independent discovery. He questioned how they could have advanced so quickly, and when he pressed for answers about data use, the company declined to explain.
The credit dispute is the visible crack. The structural break Buckmaster is pointing to is larger: when a private lab can throw 10,000 agents at a problem and produce a 166-page proof in a weekend, the entire academic publishing model — built on slow, transparent, human-verified progress — loses its footing.
The game Buckmaster says is up
Buckmaster’s accusation is specific. He alleges that after OpenAI learned of his and Alpöge’s Euler solution, the company used its AI systems to leapfrog to Navier-Stokes. He says OpenAI proposed he publish as sole author of an OpenAI paper, cutting out Alpöge — a condition he refused. Buckmaster also claims that OpenAI’s Sébastien Bubeck twice suggested removing Alpöge because he works at Anthropic, and warned Buckmaster he might ruin his career by going public. OpenAI has not directly addressed those specific claims. Bubeck later said on X that unfounded accusations could damage Buckmaster’s career, but did not directly respond to the allegations about Alpöge’s removal or the Anthropic comment.
The data-use question sits at the centre. Buckmaster asked whether OpenAI’s models were trained on or had access to his research. OpenAI spokesperson Laurance Fauconnet responded that an internal investigation found Buckmaster’s recent Codex prompts could not have influenced the Navier-Stokes proof. However, OpenAI also acknowledges in general that de-identified product data may improve its models—a distinction that Buckmaster and others find insufficient to address broader concerns about data provenance.
Terence Tao, a UCLA mathematician and Fields Medalist, calls the Euler result remarkable but warns that even rumours of progress can now trigger large-scale AI efforts aimed at scooping human projects, pushing mathematicians toward secrecy and undermining open science norms. Andreas Thom at TU Dresden has publicly demanded proof that his private ChatGPT sessions were not used in training, reflecting a broader distrust.
The Clay Mathematics Institute‘s rules add a layer of reality. To even be considered for the US$1 million prize, a solution must be published in a peer-reviewed journal, survive two years of scrutiny, and gain general acceptance in the math community. OpenAI’s Lean-verified proof remains a claim, not a recognized solution. The EU AI Act, meanwhile, requires general-purpose AI providers to publish training data summaries and respect opt-outs on text and data mining — rules that could force more transparency, though enforcement is only beginning.
The contrast with traditional mathematics is stark. Breakthroughs normally emerge from small groups over years, with proofs written for human readability. OpenAI’s effort used 10,000 agents over 88 hours to produce a 166-page machine-assisted proof and Lean code. That speed and opacity shift discovery toward compute-rich labs, weakening the role of journals and the informal priority norms mathematicians have relied on for decades.
The structural forces behind the breakdown
The dispute is not happening in a regulatory vacuum. In the US, no comprehensive federal AI-training statute exists; courts are shaping fair-use boundaries, while proposals like the CLEAR Act would force disclosure of training data sources. The EU AI Act already mandates training-data summaries and respect for opt-outs. Australia, where Buckmaster has family ties and where he gave interviews, lacks AI-specific data-training rules, relying mainly on privacy and research ethics frameworks.
The competitive landscape among AI labs adds pressure. OpenAI, Anthropic, and DeepMind are now racing to show their systems can crack frontier scientific problems. DeepMind has publicly called for shared ground rules, suggesting it sees partnership credibility as a strategic edge. Anthropic’s work with Alpöge on Euler highlights tight coupling between human insight and AI tools, while OpenAI’s Astra-driven claim showcases brute-force multi-agent reasoning.
Buckmaster’s call for academics and labs to sit down and agree on ground rules is, in effect, a demand to renegotiate the terms of engagement before the next breakthrough. The real consequence will land in twelve to eighteen months, when the Clay Mathematics Institute issues its first assessment. If Clay signals doubts, the AI labs’ proof claims lose their marketing shine. If it invites formal submissions, the review process becomes the new battleground.
Beyond the headline
The Power Behind It
The real leverage in this dispute sits with frontier AI labs that control massive compute clusters, proprietary models, and product telemetry. Those capabilities let companies decide which mathematical problems become targets and how aggressively systems chase them once a rumour appears. Universities and individual mathematicians, even at elite institutions, now depend on commercial platforms for both tools and visibility, shifting practical power over research pace, direction, and data access away from academic governance and toward corporate priorities.
The Money Trail
Even though OpenAI says it will not claim the Clay prize, the episode revolves around prestige and future revenue rather than the US$1 million itself. A verified Navier-Stokes solution serves as a marketing proof-point for Astra-class models, supporting premium pricing, enterprise deals, and investor narratives about AI’s capacity to crack unsolvable problems. The bargaining over authorship and compute offers shows how labs convert mathematical results into reputational capital, which in turn attracts talent and funding even before any prize money changes hands.
What Isn’t Being Said
Most coverage focuses on whether specific Codex prompts or chats were used, but a deeper issue is largely absent: universities and grant agencies have few safeguards for how researchers’ interactions with commercial AI tools are governed, logged, or contested. There is little open discussion of institutional responsibility when academics upload draft proofs to proprietary platforms. Bringing that missing layer into view changes the stakes from one mathematician versus one company to a systemic question about how public research infrastructure lets private labs sit inside the process of discovery itself.
Four groups facing the fallout
With the Clay Mathematics Institute’s review process now the next milestone, four groups face immediate decisions.
- Western academic mathematician using AI tools
Review your institution’s guidelines on using commercial AI tools for unpublished research. If none exist, consult national research-ethics bodies or funding-agency policies before uploading draft mathematical work to proprietary systems. The risk that private prompts become training data is real, and Buckmaster’s experience shows that even after a breakthrough, credit and control can slip away.
- US-based investor in frontier AI companies
Monitor how OpenAI and Anthropic navigate the fallout. A prolonged dispute over data provenance could invite regulatory scrutiny, especially in the EU, and erode trust among academic partners. The reputational capital at stake may influence enterprise adoption and talent acquisition, both key to long-term valuation.
- University research administrator or legal counsel
Draft clear policies on what researchers can and cannot upload to commercial AI tools. Consider requiring data-use agreements or mandating that sensitive unpublished work be processed only on institutional servers. The Buckmaster case shows that existing norms around authorship and data privacy are insufficient when a private lab can move faster than a university’s ethics committee.
- Policy professional focused on AI governance in the EU or Australia
Examine whether the EU AI Act’s training-data transparency requirements are enforceable in practice, and whether Australia’s privacy framework needs specific provisions for AI training on unpublished research. Buckmaster’s call for ground rules could become a template for mandatory consultation requirements between AI developers and academic institutions.
Explainer
- The Navier-Stokes equations describe the motion of fluids and are one of the seven Millennium Prize problems. They were formulated in the 19th century and remain unsolved in their full generality. A correct solution would have implications for weather prediction, aerodynamics, and ocean currents.
- Millennium Prize
- The Millennium Prize problems are seven mathematical challenges set by the Clay Mathematics Institute in 2000, each carrying a US$1 million reward. Only one, the Poincaré conjecture, has been solved so far. The problems were chosen to represent deep, unsolved questions that would advance mathematics significantly if cracked.
- Euler equations
- The Euler equations describe the motion of an ideal, inviscid fluid and are a simpler relative of the Navier-Stokes equations. They serve as a stepping stone for understanding fluid dynamics without viscosity. Buckmaster and Alpöge’s solution to a blowup problem for Euler was a key precursor to the Navier-Stokes dispute.
- Clay Mathematics Institute
- The Clay Mathematics Institute is a private, non-profit foundation based in Providence, Rhode Island, dedicated to increasing and disseminating mathematical knowledge. It established the Millennium Prize problems in 2000. The institute’s rules require a two-year waiting period and general acceptance before awarding any prize.
- Codex
- Codex is an OpenAI system that translates natural language into code, used by researchers to draft and test mathematical proofs. It is distinct from ChatGPT, focusing specifically on programming tasks. Buckmaster used Codex prompts in the months before the dispute, which OpenAI says did not influence its Navier-Stokes model.
- Lean
- Lean is a proof assistant and programming language developed by Microsoft Research that allows mathematicians to write and verify proofs in a formal, computer-checkable way. OpenAI’s Navier-Stokes proof was accompanied by Lean code to demonstrate its formal correctness. The use of Lean is becoming standard in AI-driven mathematics to provide machine-verifiable evidence.
RELATED STORIES
