WASHINGTON — The problem has remained open since 1845. For 181 years, the world’s leading mathematicians have been unable to prove whether the equations describing fluid motion remain smooth indefinitely or eventually develop infinite turbulence. On Tuesday, OpenAI said its artificial-intelligence system had solved the problem. By the time the announcement was published, however, a dispute over credit had already emerged.
The Navier–Stokes existence and smoothness problem is one of seven Millennium Prize Problems identified by the Clay Mathematics Institute in 2000, each carrying a $1 million award. Six remain unsolved. On September 8, OpenAI published a 166-page proof and Lean formalization arguing that a smooth external force can cause a three-dimensional fluid to develop unbounded velocity in finite time while its kinetic energy remains bounded. The result supports the “blow-up” scenario—one of the two possible resolutions to the problem.
According to the company, the proof was produced not by mathematicians but by an internal model significantly more capable than GPT-6 Astra, the frontier model OpenAI released six days earlier. Ten thousand AI agents worked concurrently for 88 hours, generating approximately 2.7 million messages at an average of 48,000 tokens each. OpenAI disclosed the cost as approximately $6.5 million. GPT-6 Astra’s specific role was the Lean formalization, which expressed the proof in a formal language that enables computers to verify each logical step independently.
The result has implications beyond pure mathematics. Models of turbulence used in weather forecasting, aircraft design, and fluid engineering generally assume that the Navier–Stokes equations remain smooth. A formal proof that smooth initial conditions can produce a finite-time singularity introduces a limitation to those assumptions. Whether that limitation affects practical engineering will depend on the energy regimes described by the proof—a question applied scientists are likely to examine in the coming months. The theoretical landscape has nonetheless changed: one of the two possible answers to the smoothness question now has a machine-verified result behind it, as detailed in the company’s published proof.
The Clay Institute has not completed its formal review. Its prize process requires eventual publication in a peer-reviewed journal, followed by a two-year waiting period before formal consideration. The $1 million prize has not been awarded, and no outside institution has confirmed that it will be.
What preceded the announcement is a more immediate story.
The rumor nevertheless pointed to genuine research. New York University mathematics professor Tristan Buckmaster and Anthropic researcher Levent Alpöge had spent nearly a year studying the forced Euler equations, a related but distinct problem, and achieved a breakthrough on August 15.
Buckmaster published a four-page statement on September 7 describing what happened when OpenAI contacted both researchers to discuss a joint announcement. According to Buckmaster’s account, Bubeck offered two paths: Buckmaster and Alpöge could publish their work first, with OpenAI following the next day, or Buckmaster could publish alone and claim the prize, but only if he removed Alpöge’s name from the paper. The stated reason: OpenAI did not want the announcement to include an Anthropic employee as co-author. Buckmaster refused. TechCrunch reported the full account of the negotiations.
OpenAI has denied that its researchers saw Buckmaster and Alpöge’s unpublished work before OpenAI’s own release. The company also denied inspecting private user data, while acknowledging, without detail, that de-identified platform usage could have informed its internal model’s training. It has not directly responded to Buckmaster’s specific account of the pre-announcement negotiations.
The dispute arrived in the middle of a remarkable week for the sector. GPT-6 Astra, which OpenAI describes as the world’s most intelligent and aligned model, scored 98 percent on the FrontierMath Tier 4 benchmark and is already in commercial release. The unnamed internal model that produced the Navier-Stokes proof in 88 hours is, in OpenAI’s own characterization, a generation beyond Astra. Mistral AI closed the largest equity round in European tech history the same week, reinforcing a pattern in which capability gains and capital pile up faster than the institutions designed to evaluate either.
Fields medalist Terence Tao, who has engaged extensively with AI-assisted mathematics, wrote on his website that the announcement represents “a genuine and significant result, though one whose boundary conditions and hypotheses need careful independent examination before any prize conversation begins.” He did not address the credit dispute.
The deeper question raised by the dispute concerns not these researchers specifically, but the meaning of intellectual priority when artificial intelligence compresses the timeline for solving century-old mathematical problems from years to days.
Buckmaster and Alpöge spent nearly a year working on the forced Euler equations, making extensive use of Claude and Codex. Once deployed on the related Navier–Stokes problem, OpenAI’s internal model produced a comparable result in less than four days. The researchers were operating at the frontier of what human–AI collaboration can currently achieve. The machine, given greater scale, moved beyond that frontier without stopping to consider what it meant to be first.
That acceleration is the actual news. The credit dispute is a symptom of it.
The Clay Mathematics Institute’s review will take time. The Lean verification makes the proof’s logical structure checkable by anyone with the mathematical preparation to read it, and scrutiny is already underway, as Engadget reported. What Tao’s comment implies, and what Buckmaster’s four-page account makes concrete, is that the mathematics may be confirmed before the story of how it was produced is fully resolved.
