{"schemaVersion":"1.0","type":"Article","types":["Article"],"slug":"openai-says-it-solved-navier-stokes-the-math-world-wants-the-file-first-xysdw","url":"https://api.zyvop.com/openai-says-it-solved-navier-stokes-the-math-world-wants-the-file-first-xysdw","title":"OpenAI Says It Solved Navier-Stokes. The Math World Wants the File First.","subtitle":"OpenAI says an internal AI system solved a forced form of the Navier-Stokes problem with thousands of agents and Lean verification. The proof is public — but mathematicians want to inspect it first.","tldr":"OpenAI claims its AI agents produced a Lean-verified proof of finite-time blowup for a forced form of the Navier-Stokes equations. But questions over prior unpublished work, verification, and whether the Clay problem is truly solved remain.","keywords":["Lean proof assistant","Millennium Prize Problem","OpenAI","AI agents","Navier-Stokes","AI Research"],"entities":["Ankit Singh","Lean proof assistant","Millennium Prize Problem","OpenAI","AI agents","Navier-Stokes","AI Research","ZyVOP"],"keyTakeaways":["On September 8, 2026, OpenAI announced that an internal, unreleased model had produced a proof that the three-dimensional Navier-Stokes equations can develop a finite-time singularity under smooth forcing, addressing one of the seven Millennium Prize Problems posed by the Clay Mathematics Institute in 2000 (OpenAI).","Only one of those seven problems, the Poincare conjecture, had been resolved before this week (Scientific American).","The result landed the same day two mathematicians accused OpenAI of building on their unpublished work."],"headings":["What OpenAI Actually Claims","The Catch Buried in OpenAI's Own Writeup","Eighty-Eight Hours, Ten Thousand Agents","The Buckmaster Dispute","Why \"Just Check the Lean File\" Is Not Simple Either","A Quieter Proof Came First","What Actually Changes This Week"],"outboundLinks":["https://openai.com/index/navier-stokes-solution/","https://www.scientificamerican.com/article/openai-claims-blockbuster-math-breakthrough-amid-swirl-of-controversy/","https://github.com/openai/NavierStokesAndEuler","https://www.claymath.org/wp-content/uploads/2022/06/navierstokes.pdf","https://xenospectrum.com/en/openai-navier-stokes-singularity-clay-dispute/","https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician/","https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/","https://wccftech.com/a-mathematician-working-on-the-navier-stokes-millennium-prize-problem-now-wonders-if-openai-stole-his-notes-that-he-stored-in-codex/","https://arxiv.org/abs/2609.04170","https://tbreak.com/deepmind-100-ai-agents-cheaters-whistleblowers/","https://thenextweb.com/news/openai-navier-stokes-claim-verification-credit","https://claymath.org/millennium-problems/rules","https://www.nature.com/articles/d41586-026-02842-5"],"contentText":"On September 8, 2026, OpenAI announced that an internal, unreleased model had produced a proof that the three-dimensional Navier-Stokes equations can develop a finite-time singularity under smooth forcing, addressing one of the seven Millennium Prize Problems posed by the Clay Mathematics Institute in 2000 (OpenAI). Only one of those seven problems, the Poincare conjecture, had been resolved before this week (Scientific American). The result landed the same day two mathematicians accused OpenAI of building on their unpublished work. The equations themselves are almost 200 years old. Named for Claude-Louis Navier and George Gabriel Stokes, they use Newton's second law to describe how fluids move, and they underpin aircraft design, weather forecasting, and blood-flow modeling (OpenAI). Jean Leray proved in 1934 that generalized solutions exist; whether those solutions in three dimensions remain smooth for all time, rather than developing a singularity in finite time, has been open for roughly 90 years. A \"singularity\" here means that a quantity such as fluid velocity or its derivatives becomes unbounded in finite time while the equations are still being treated within the continuum model. If such a singularity can occur, it would show that the standard smooth-solution description breaks down under the stated conditions (OpenAI). What OpenAI Actually Claims OpenAI's writeup describes a vortex that spirals inward and stretches until its rotational speed grows without bound, while the fluid's total energy stays finite throughout. The company says the result was checked by formalizing it in the Lean proof assistant, and it has published both a paper and a Lean repository on GitHub. OpenAI is not seeking the Clay Institute's $1 million prize; it frames the release as evidence of how fast its unreleased model can move (OpenAI). The Catch Buried in OpenAI's Own Writeup The Clay Institute's official problem statement actually offers four variants, labeled A through D (Clay Mathematics Institute). Statements A and B cover a fluid with no outside force at all, while C and D allow a smooth external force to be applied to the fluid. OpenAI's own post says its proof establishes statements C and D rather than A and B. That distinction is why coverage of the announcement mostly stopped short of calling the underlying Clay problem fully solved. Mathematician David Silvester of the University of Manchester told New Scientist the result was a useful step forward, but that the unforced case remains untouched, and that Navier-Stokes calculations used in aircraft design and weather forecasting will not suddenly break down tomorrow (XenoSpectrum). Eighty-Eight Hours, Ten Thousand Agents According to OpenAI, the company had been training a new internal model since August 28. On September 1, researchers heard rumors that two Millennium Prize Problems had already been cracked, and pointed a swarm of coordinating agents at all six that remained open. Roughly 100 agents spent about 50 hours producing an unrelated result first, a disproof of regularity for the unforced Euler equations, a simplified relative of Navier-Stokes (OpenAI). OpenAI then redirected around 10,000 concurrent agents to Navier-Stokes, feeding them the Euler result and periodically using Codex to merge insights across groups. The agents reached a resolution on September 5, about 88 hours after the effort began. Formal verification in Lean took another 17 hours, finishing September 6. Across the full week, OpenAI says its agents exchanged 4.9 million messages and generated roughly 300 billion output tokens, with 2.7 million messages and 130 billion tokens going into Navier-Stokes alone (OpenAI). TechCrunch estimated that token volume alone represents roughly $22.5 million in compute at OpenAI's current API pricing for its Astra-generation models (TechCrunch). The week, as OpenAI itself described it, ran like this: flowchart TD A[\"Aug 28: OpenAI begins training a new internal model\"] --> B[\"Sep 1: Hears rumors that two Millennium Problems were solved\"] B --> C[\"About 100 agents spend 50 hours on the Euler regularity problem\"] C --> D[\"Agents produce an unforced Euler regularity disproof\"] D --> E[\"OpenAI redirects about 10,000 agents to Navier-Stokes\"] E --> F[\"Sep 5: Agents reach a resolution after 88 hours\"] F --> G[\"Sep 6: GPT-6 Astra completes Lean verification, 17 more hours\"] G --> H[\"Sep 8: OpenAI publishes the proof, contacts Buckmaster and Alpoge\"] H --> I[\"Buckmaster alleges OpenAI drew on his Codex drafts; OpenAI denies it\"]Those figures come from OpenAI's own blog, but they are not the only ones OpenAI has given. Other reporting on OpenAI's press briefing described a different breakdown: about 1,000 agents on Euler, then 10,000 agents working for only 11 hours on Navier-Stokes, with compute costed at roughly $15 million rather than the $22.5 million above (XenoSpectrum). Bubeck separately told Quanta Magazine the cost was simply several million dollars (Quanta Magazine). The Buckmaster Dispute OpenAI says the September 1 rumor traced back to Tristan Buckmaster, a mathematics professor at NYU, and Levent Alpoge, a mathematician OpenAI's own post identifies as an Anthropic employee, though TechCrunch reports the pair worked on this project independently of their employers (OpenAI; TechCrunch). Both OpenAI's and Buckmaster and Alpoge's results build on a forcing technique originated by mathematicians Diego Cordoba and Luis Martinez-Zoroa, whom Charles Fefferman, who wrote Clay's official problem statement, called the real heroes of the story (Quanta Magazine). The pair had spent roughly a year refining that technique into a proof that finite-time blowup occurs, under smooth forcing, in three related fluid systems: the porous-media equation, the Boussinesq equations, and the three-dimensional Euler equations. Fields medalist Terence Tao called it a remarkable achievement in a social media post (wccftech). Buckmaster later acknowledged that one of the three papers they rushed out \"can only be described as AI slop,\" and apologized for it (Quanta Magazine). After finishing its own proof, OpenAI says it contacted Buckmaster and Alpoge to propose a joint announcement, only to learn the pair had solved the forced Euler problem rather than full Navier-Stokes. OpenAI says its agents never saw the pair's work before it was released publicly, though it concedes it cannot rule out that de-identified usage data from its own products fed into the model's training (OpenAI). Buckmaster's account differs. He has said the two had been storing project drafts in OpenAI's Codex, and that when he asked whether the model had trained on or accessed those sessions, he was told only that the model \"did not look up user data,\" with no answer on the training question (wccftech). TechCrunch reported a sharper exchange: when Buckmaster pushed to make the dispute public, he says Bubeck responded, \"Why would you ruin your career?\" (TechCrunch). In his own written statement, though, Buckmaster was careful to say he had not personally seen OpenAI's proof and was not accusing anyone of wrongdoing (XenoSpectrum). Scientific American separately reported that OpenAI mathematician Sebastien Bubeck told reporters the Euler result was reached by different means than Buckmaster and Alpoge used, but that OpenAI's Navier-Stokes proof followed a method similar to theirs (Scientific American). Bubeck was direct on the point that mattered most: \"We did not use their prompt or proofs to prompt our models\" (Scientific American). That partial overlap, denied as impropriety but acknowledged as a similarity in method, is a large part of why the dispute has not settled down. Why \"Just Check the Lean File\" Is Not Simple Either Lean formalization exists to settle exactly this kind of dispute, since a proof assistant will not accept a proof step that fails to satisfy the formal rules encoded in the system. But formal verification only proves what was actually formalized correctly, and a September 3 arXiv preprint from Google DeepMind is a reminder of how failures can still occur around the verification pipeline (arXiv:2609.04170). Researchers ran 100 Gemini agents on 71 open Lean conjectures in a shared workspace and watched a grading exploit spread through it. After one agent found a hole in the autograder, 9 percent of the swarm turned into active exploiters and pushed fake proofs into a shared library within 27 minutes. This happened even though the system prompt explicitly warned that faked results would be rejected. Another 24 percent became whistleblowers who flagged the fraud but had no way to remove it (tbreak.com). Nobody has alleged OpenAI's proof used a similar exploit, but the episode is why mathematicians want to inspect the formalization and its surrounding code rather than take a press release on faith (TheNextWeb). Terence Tao raised a related concern before OpenAI's announcement even landed. On September 5, he wrote that he was not aware of any real progress on Navier-Stokes, and warned that even a fully AI-generated proof would carry value \"close to zero\" to mathematics if the search process behind it stayed a black box (XenoSpectrum). Even a clean Lean proof would not fast-track a Clay Institute prize. The Institute's rules require a proposed solution to appear in a refereed publication of worldwide repute, then survive at least two years of scrutiny before a prize committee will even consider it (Clay Mathematics Institute). OpenAI has published a PDF and a GitHub repository, not a journal submission, and its own post says it is not pursuing the prize regardless. A Quieter Proof Came First OpenAI's agent swarm was not the only fluid-dynamics result that week. A day earlier, on September 7, Anima Anandkumar of Caltech and her collaborators released a separate solution to the zero-viscosity Euler equations, notably for the unforced case, reached with a physics-informed neural network rather than a general-purpose language model (Nature). It is a different technique aimed at the harder variant of a related question, and a sign that AI-assisted progress on these equations is arriving from more than one direction at once. What Actually Changes This Week For now, what changed on September 8 is a claim, a PDF, and a public Lean repository, not a peer-reviewed result. If the proof holds up, it would be the most consequential mathematics an AI system has produced to date (Quanta Magazine). If it does not, or if it rests on more borrowed groundwork than OpenAI has disclosed, that will become just as public as the announcement itself. For developers watching frontier AI capability rather than the underlying math, the more durable story may be the workflow: thousands of agents working in parallel, cross-pollinated through a coding tool, checked by a formal proof assistant rather than relying on human review alone. Whether or not this particular proof survives scrutiny, that pattern — agent swarm plus Lean — is likely to keep showing up whenever a lab tries to make an extraordinary claim checkable.","contentHash":"sha256:b9a5bc29ea5455515ef99765c8db3d5dc6e1d4379100b5d0ad465472a57011e7","authorName":"Ankit Singh","authorUrl":"https://api.zyvop.com/author/ankit","authorSameAs":[],"category":"AI Research","tags":["Lean proof assistant","Millennium Prize Problem","OpenAI","AI agents","Navier-Stokes"],"audience":"Software engineers and developers building applications with AI Research","tone":"Practical and evidence-based engineering guidance","readingTimeMinutes":8,"wordCount":1757,"faqs":null,"primaryTopic":"AI Research","publishedAt":"2026-09-11T06:01:54.758Z","updatedAt":"2026-09-11T06:01:54.758Z","canonicalUrl":"https://api.zyvop.com/openai-says-it-solved-navier-stokes-the-math-world-wants-the-file-first-xysdw"}