When Machines Claimed a Millennium Prize Problem

At noon on 8 September 2026, OpenAI announced that one of the most famous open problems in mathematics had fallen after nearly ninety years of resistance. The company said its systems had resolved the Navier-Stokes existence and smoothness question, a Millennium Prize challenge carrying a one-million-dollar award . The result describes a smooth flow, pushed by a smooth external force, that still develops a finite-time singularity, covering the cases labeled (C) and (D) by Charles Fefferman. A 166-page write-up arrived together with a Lean machine-checked formalization, produced by an internal model described as markedly stronger than GPT-6 Astra. Coverage of the noon announcement and the prize context was summarized by OpenAI in a same-day technical briefing with full derivation logs.

The underlying mathematics concerns the equations that govern viscous fluid motion, where velocity and pressure evolve under friction, inertia, and forcing. A singularity, sometimes pictured as a tightening spaghetti-like vortex, would mean that perfectly smooth starting conditions still collapse into an infinitely sharp spike. Viscosity was long expected to prevent such a breakdown by smoothing sharp gradients, so a finite-time blow-up would overturn deep intuition about fluids. The modern regularity question descends from Jean Leray, who in 1934 constructed global weak solutions and left open whether smooth data always stay smooth. Clay Institute formulations turned that legacy into one of seven official Millennium challenges with a shared multi-million purse.

Ten Thousand Agents and Eighty-Eight Hours

The claimed computation began with an internal model trained since 28 August, redirected on 1 September toward every open Millennium problem after rumors that two such problems had fallen. Roughly 10,000 concurrent agents ran for 88 hours, from 1 to 5 September, exchanging 2.7 million messages and emitting about 130 billion output tokens at a power bill in the millions, a figure confirmed by researcher Noam Brown. The fleet first rehearsed on the unforced Euler variant with about 100 agents in roughly 50 hours, then shifted its weight to Navier-Stokes while groups cross-pollinated through Codex, adding 17 more hours of Lean checking with GPT-6 Astra. OpenAI said the same run would cost an outside customer about $15 million at list prices, a striking measure of scale reported in detail by NewScientist in a September analysis of the compute campaign.

Parallel to that industrial push, NYU Courant faculty member Tristan Buckmaster and Anthropic mathematician Levent Alpoge had pursued the same frontier quietly for almost a year. The pair describe their effort as a strictly personal collaboration rather than an institutional alliance, conducted with models from both companies. Every draft lived inside Codex sessions, which later became central to the dispute about what the large fleet might have absorbed. Their focus was forced variants of related fluid models, a path they believed could reach further than the celebrated unforced constructions. Secrecy was deliberate, since they wanted a human-readable paper before any public announcement.

By 15 August the duo had obtained smooth forced results for the Euler, IPM, and Boussinesq systems, verified in Lean about a week later after careful checking. They held back publication because they wanted an exposition that working mathematicians could read, check, and teach, rather than a bare formal artifact. Their scaffolding drew on the forced and hypodissipative advances of Diego Cordoba, Luis Martinez-Zoroa, and Fan Zheng, plus the 2013 numerical evidence of Luo and Hou that first suggested a plausible blow-up route. Standard background on the Clay Millennium challenges and the history of the regularity question is collected on Wikipedia in a long overview of the Millennium Prize problem list.

The public spark came from Alpoge, whose 31 August note titled Augustus Mirabilis celebrated striking progress and was misread online as a claim that two Millennium problems had already been solved at Anthropic. The viral rumor traveled fast across social platforms and group chats, prompting OpenAI leadership to launch its own full-scale assault on 1 September. What began as an exuberant laboratory memo thus turned into a race neither side had planned, with the calendar suddenly compressed to days. The irony is sharp: a message meant to mark a wonderful August pushed a rival lab into overdrive within hours.

Letters, Meetings, and a Midnight Statement

Warned of a possible leak, the pair wrote to OpenAI separately: Alpoge late on 2 September and Buckmaster on 3 September to researcher Sebastien Bubeck. On 6 September Bubeck replied that the effort used very little human input and offered to share everything including human directives, followed by an afternoon meeting that Alpoge did not attend. Buckmaster recognized his own Cordoba and Martinez-Zoroa program in the machine route, which sat uneasily with claims of minimal guidance and a large team reportedly containing no Navier-Stokes specialist. Direct file access was refused, while later statements conceded the use of de-identified derivative data in training. The exchange left the central question unresolved: inspiration, coincidence, or ingestion.

Just before midnight on 7 September, Buckmaster posted a four-page public statement laying out the chronology from his side, roughly twelve hours before the scheduled corporate announcement. OpenAI then unveiled its Navier-Stokes result at noon on 8 September as planned, complete with manuscript, formalization, and cost figures. Discussion erupted across Hacker News, Reddit, and X within hours, mixing excitement about the mathematics with unease about provenance. Buckmaster finally left his office around 02:30 that morning, after a night spent answering messages from colleagues worldwide. The twelve-hour gap between the two publications fixed the dispute in public view.

The official account shifted repeatedly over the following days, each version narrowing or widening the window of possible contact. An 8 September morning line said nothing was seen before publication and that de-identified derivative material could not be ruled out, followed by remarks that nobody had called at all. By that afternoon the emphasis moved to visible differences between the proofs, especially the forced-versus-unforced distinction in Euler, then on 9 September to a categorical denial covering the prior two months. A late 10 September update was followed on 13 September by a firm cutoff: nothing entered after 3 July could have mattered, with no comment on heavy Codex use before that date . Observers noted that each revision answered the previous criticism while opening a new one.

Behind the statements lay a sharper private quarrel over authorship and money. According to Buckmaster, OpenAI floated options including announcing Euler first, or rewriting its Navier-Stokes proof with him as sole or lead author alongside a $1 million payment to the pair as the people closest to the problem. One condition attached was that Alpoge be removed from authorship, a term Buckmaster says he rejected outright. The correspondence turned blunt, with a warning that politeness was optional, for which Bubeck later apologized over tone. Whatever the intent, the episode showed how credit and compensation can collide when automated discovery meets human priority claims.

What Mathematicians Say Comes Next

Terence Tao called the duo achievement remarkable and saw no principled barrier to extending it toward Navier-Stokes, while warning that industrial-scale proof mining risks strip-mining the field. In an interview he compared the practice to looting an archaeological site with an excavator and said 2026 had split answers from understanding, likening the situation to watching only the opening and closing scenes of a film. On 11 September, 28 Fields medalists signed a statement titled A Severe Misalignment of AI in Mathematics urging slower and more transparent norms. Sociologist Michael Harris added that without profit such lavish runs cannot persist, yet decision-makers may still conclude that humans are expendable, a concern echoed in an IBM discussion of scientist-machine collaboration in mathematical discovery.

The prize machinery itself moves far more slowly than press cycles, requiring publication in a qualifying journal, a two-year waiting period, and broad acceptance by the mathematical community. President Ulrike Bridson has described the process as deliberately slow and absolutely rigorous, a design meant to survive exactly this kind of commotion. OpenAI says it does not intend to claim the award, leaving the forced-versus-unforced eligibility question to mathematicians rather than to marketing. Prize rules and the history of the seven challenges are maintained by ClayMath in its official Millennium archive with dates and citations. A plain-English recap of the announcement sequence and the shifting statements was published by TheNextWeb in a mid-September explainer with links to primary documents.

On 6 October the project widened into a bulk release: 722 manuscripts, 372 families of results, and hundreds of open questions, with a public GitHub repository, Lean formalizations, ten reasoning summaries, and compute estimates averaging three hours of ChatGPT Pro thinking per result. Attached AGMAI proposals call for rapid publication, disclosure of models, directives, and costs, and a pledge not to turn proofs into marketing instruments. Host Dagogo Altraide framed the arc as a Deep Blue versus Kasparov moment, a genuine landmark for pure mathematics with limited near-term practical payoff since working fluid formulas already perform well. A compact video-centered recap of the October data drop and its reception was posted by TheVerge in a same-week roundup of frontier-model science claims.