10,000 AI Agents and the Navier-Stokes Breakthrough
OpenAI says roughly 10,000 agents produced a Lean-verified answer to a Millennium Prize problem in 88 hours. The proof, the credit dispute, and the Fields Medalists' warning.
The Claim: A Millennium Problem Falls to a Swarm
On September 8, 2026, OpenAI announced that roughly 10,000 autonomous agents, running on an internal model it describes as significantly more capable than the GPT-6 Astra released a week earlier, produced a result on the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000, open for roughly 90 years. The run consumed about 88 hours, 2.7 million messages between agents, and roughly 130 billion output tokens, after which GPT-6 Astra spent another 17 hours writing a machine-checkable Lean formalization, published on GitHub.
What was actually proved matters. The result shows that a smooth three-dimensional fluid initially at rest can, under a smooth forcing term applied throughout, develop a singularity — a point where velocity climbs without bound — in finite time, while the fluid's energy remains finite. OpenAI itself frames this as establishing statements C and D of the official Millennium Prize formulation rather than the classic unforced problem, and on September 10 it published the full technical write-up alongside the Lean code. The company also stated it will not claim the 1 million dollar prize.
The Credit Dispute That Opened the Same Day
Twelve hours before OpenAI's announcement, NYU mathematician Tristan Buckmaster had announced that he and Levent Alpöge, a mathematician at Anthropic, resolved several closely related problems after nearly a year of working with public models including OpenAI's Codex and Anthropic's Claude. Buckmaster alleges OpenAI learned of their progress and raced ahead using its compute advantage; he says OpenAI's Sébastien Bubeck asked him to drop Alpöge's name because Alpöge works for a competitor, and that when he pushed to go public he was asked, in his account, "Why would you ruin your career?"
OpenAI's chief research officer Mark Chen denied that any agent or employee accessed the pair's work. The company says it began on September 1 after hearing a rumor, and that while it cannot rule out that de-identified usage data helped improve its models, no specific user data was accessed. Buckmaster's side says they had a Lean-verified proof for the Euler equations by August 22. Both efforts built on an approach opened by Diego Córdoba and Luis Martínez-Zoroa. Each contested point remains an allegation met by a denial, with no third-party resolution.
Verification Replaces Peer Review — At a Price
The Lean formalization is the structural novelty: anyone with a computer can machine-check the proof against its stated premises, without trusting the author or the institution. Clay's own rules, however, are deliberately slow — a prize requires peer-reviewed publication and two years of community validation, and only one Millennium Problem, the Poincaré conjecture, has ever been settled. The computing cost is its own story: OpenAI described the run as emphatically in the millions of dollars, roughly 1,000 times what it spent on earlier mathematical results, according to researchers' comments to the press.
The resource gap worries mathematicians more than the speed. Two researchers with public models produced a partial proof over a year; a lab with an internal model and eight-figure compute produced a full proof in days. Very few academic mathematicians will ever command compute at that scale.
Fields Medalists Push Back
On September 11, 25 Fields Medalists including Terence Tao published a joint declaration, "A Severe Misalignment of AI in Mathematics," arguing that the field's most precious resources are students and ideas, and that AI rapidly producing propositions without careful verification and discussion risks severing the chain of human knowledge transmission. Tao had argued earlier that the value of a hard problem lies in the field-wide progress generated by human effort, and that premature AI solutions without transparency into the process can contaminate that progress.
What to Watch
Three threads remain open: whether the mathematical community's independent review confirms the proof's assumptions and dependencies, whether the Buckmaster dispute produces any accountability mechanism for "vendor data feedback" — researchers' intermediate work feeding back into model vendors' training — and whether Lean-verified delivery becomes the standard contract for AI-generated mathematics. The problem may have fallen, but the institutions around it are only beginning to adapt.
Sources: The Century Report | AI之上 OpenAI Navier-Stokes Proof Published