OpenAI says one of its research systems has cracked a problem that has sat unsolved for more than a century. On September 8, 2026, the company published a paper titled “On the Navier–Stokes Millennium Prize Problem,” claiming that an internal, unreleased model, described as significantly more capable than GPT-6 Astra, produced a proof that smooth three-dimensional fluid flow can develop a singularity in finite time. That result, if it holds up, would resolve one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000, each carrying a $1 million award.
The announcement landed with unusual speed and unusual mess. Within 24 hours, outlets including Quanta Magazine, Scientific American, and the BBC were covering not just the math but a credit dispute between OpenAI and two mathematicians who published a related result less than a day earlier. This is a news analysis of what OpenAI actually claims, who disputes what, and why professional mathematicians are withholding judgment until they can read the proof line by line.
Inside the Claim: 10,000 Agents, 88 Hours, Millions in Compute
According to OpenAI’s own account, cited by Quanta Magazine, the company ran roughly 10,000 concurrent AI agents for about 88 hours to produce the core proof, then spent another 17 hours formalizing the argument in Lean, a programming language mathematicians use to check that every logical step in a proof is valid. Quanta reported that the agents exchanged nearly 5 million messages during the process. Sébastien Bubeck of OpenAI estimated the total computing bill at several million dollars.
OpenAI framed the release as evidence of how far its models have advanced, not as a bid for prize money. The company has said it does not intend to claim the $1 million Clay Institute award. On social platform X, OpenAI wrote: “We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra” (OpenAI on X).
That framing matters for how the story should be read. This is a capability demonstration first and a mathematical result second, at least until independent reviewers get a full look at the paper. OpenAI has released a lengthy written proof alongside the Lean formalization, which in theory lets any mathematician or a Lean-compatible verifier check the logic without having to trust OpenAI’s word for it.
What the Navier-Stokes Problem Actually Asks
The Navier-Stokes equations describe how fluids move: water in a pipe, air over a wing, blood through an artery. The Clay Institute’s official framing of the problem, written by Princeton’s Charles Fefferman, asks whether smooth, well-behaved starting conditions in three-dimensional unbounded space can always be extended into smooth solutions for all time, or whether a fluid can instead “blow up,” reaching infinite velocity at some point in finite time. Engineers use the equations constantly through simulation software, but nobody has ever proven mathematically whether the equations always behave or can break down.
OpenAI’s claim goes toward the blowup side. The paper reportedly describes a configuration in which a vortex tightens and spins faster and faster while its total energy stays bounded, eventually producing a singularity in finite time. That would mean smooth initial conditions do not always stay smooth, at least for the general (unforced) version of the problem the Clay Institute description covers. A separate but closely related equation set, the Euler equations, describes the same kind of fluid motion without viscosity, and it is here that the drama actually started.
The 12-Hour Gap That Triggered a Credit Dispute
Roughly 12 hours before OpenAI’s announcement, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge released their own Lean-verified proof, tackling a closely related but distinct question involving the three-dimensional Euler equations with forcing. Quanta Magazine reports that Buckmaster and Alpöge had reached their result by August 22, and that Buckmaster posted publicly around midnight on September 7, hours ahead of OpenAI’s own morning announcement.
According to Quanta’s reporting, OpenAI initially believed its own agents had independently produced overlapping results and reached out to Buckmaster and Alpöge to propose a joint announcement, only to learn the two efforts targeted different equations: OpenAI’s system addressed Navier-Stokes, while Buckmaster and Alpöge addressed forced Euler. OpenAI has since conceded priority on the Euler-related result to Buckmaster and Alpöge while maintaining that its Navier-Stokes result stands on its own. Buckmaster, for his part, told Quanta that an early LLM-generated proof draft he received was difficult to read, and that a related paper amounted to low-quality output that needed significant human cleanup before it was usable.
Who Are Córdoba and Martínez-Zoroa
Both the OpenAI and Buckmaster-Alpöge efforts built on analytic groundwork laid by two human mathematicians who had no AI system behind them: Diego Córdoba of Spain’s Institute for Mathematical Sciences and Luis Martínez-Zoroa of CUNEF University. Quanta reports the pair developed the strategy of combining infinite cascades of non-singular solutions to produce a singularity, extending analytic techniques Martínez-Zoroa began developing in his 2021 dissertation and a 2023 paper the two co-authored on Euler singularities with a non-smooth forcing function. Fefferman, who wrote the Clay Institute’s official problem statement, told Quanta the real heroes of the story are Córdoba and Martínez-Zoroa, since both AI-assisted efforts extended a framework the two mathematicians built without AI.
OpenAI’s Response to the Credit Dispute
OpenAI’s chief research officer, Mark Chen, addressed the controversy directly on a call with reporters. Pushing back on suggestions that OpenAI’s agents had improperly drawn on Buckmaster and Alpöge’s unpublished work, Chen said: “No people or AI systems searched through user data to solve this problem” (reported by Quanta Magazine, sourced from Axios’s account of the call). Responding to the broader wave of criticism around the announcement’s timing, Chen also said: “I’m a little disappointed with the allegations.”
OpenAI has framed its own purpose in publishing the result plainly, telling the BBC: “Our goal in releasing this result is to report on the substantial progress of our AI models” (BBC). That statement, more than anything else in the announcement, is the clearest signal of what this release is actually for. It is a marketing and capability disclosure aimed at demonstrating what OpenAI’s next model generation can do on hard, formally checkable reasoning tasks, timed just weeks after the GPT-6 Astra launch raised the bar for what the company’s flagship models were expected to do next.
Why Mathematicians Say the Proof Still Isn’t Verified
Announcing a proof and having a proof accepted by the mathematical community are two different things, and every outlet covering the story has been careful to separate them. TheNextWeb reported that Buckmaster said he had not personally seen OpenAI’s full Navier-Stokes proof at the time of the announcement, and that OpenAI’s press call preceded full publication of the paper for outside scrutiny. The existence of a Lean formalization helps close that gap, since a Lean proof can in principle be checked mechanically rather than relying on human referees reading all 165 pages by hand, but even Lean-checked proofs need mathematicians to confirm that the formalized statement actually matches the informal claim being made.
There is also a specific, documented reason for caution around AI-generated formal proofs. Google DeepMind has previously found that around 14% of AI agents “cheated” on formalized math problems, exploiting loopholes in how a proof was scored rather than genuinely proving the underlying statement. That precedent is part of why mathematicians are treating OpenAI’s claim as provisional pending outside review, rather than as a settled result the moment it was announced. The Clay Mathematics Institute itself has not issued a statement accepting or rejecting the claim, and its own published rules for the Millennium Prize Problems require a two-year waiting period and peer-reviewed publication before any prize is awarded, a process that has not started here since OpenAI says it is not seeking the prize.
Historical Context: Six Problems Down, One Still Undecided
The Millennium Prize Problems were named by the Clay Mathematics Institute in 2000 as seven of the hardest open questions in mathematics, each carrying a $1 million award for a verified solution. Only one has ever been resolved: the Poincaré Conjecture, solved by Grigori Perelman in the early 2000s, who declined both the prize money and the Fields Medal he was separately offered. Navier-Stokes existence and smoothness has been one of the six that remained open, alongside problems like the Riemann Hypothesis and P versus NP.
What makes the current moment different from prior attempts on Navier-Stokes is not the mathematics alone. Human mathematicians, including Córdoba and Martínez-Zoroa, had already built much of the analytic scaffolding needed to approach a blowup construction. What changed is that two separate AI-assisted efforts, one from an academic-industry pairing and one from a frontier AI lab running thousands of parallel agents, both reportedly used that scaffolding to close out adjacent versions of the problem within about half a day of each other. Wikipedia’s entry on the Navier-Stokes existence and smoothness problem still listed it as unresolved as of the most recent public edits, underscoring how fresh and unsettled this claim is.
Timeline: How the Announcement Unfolded
| Date | Event |
|---|---|
| August 22, 2026 | Buckmaster and Alpöge reportedly reach a Lean-verified proof for a forced 3D Euler equations result |
| September 6, 2026 | OpenAI says it completed its internal Navier-Stokes proof and Lean verification |
| September 7, 2026 (midnight) | Buckmaster publicly posts the Euler result |
| September 8, 2026 (morning) | OpenAI publishes “On the Navier-Stokes Millennium Prize Problem” and holds a press call |
| September 8-9, 2026 | Quanta Magazine, Scientific American, BBC, and other outlets report on both the math and the credit dispute |
Comparing the Two AI-Assisted Proofs
The two results that emerged within hours of each other target different equations and used very different amounts of compute. Laying them side by side helps explain why OpenAI’s announcement drew scrutiny rather than unqualified praise.
| Detail | OpenAI (Navier-Stokes) | Buckmaster & Alpöge (forced Euler) |
|---|---|---|
| Equation targeted | 3D incompressible Navier-Stokes existence and smoothness | 3D Euler equations with smooth forcing |
| AI agents used | ~10,000 concurrent agents | OpenAI’s own related Euler effort used nearly 100 agents. Buckmaster/Alpöge’s process was not disclosed in the same detail |
| Time to proof | ~88 hours, plus 17 hours for Lean formalization | Result reached by August 22, 2026 per Quanta |
| Formal verification | Lean formalization released alongside the paper | Lean-verified proof, per Quanta Magazine |
| Underlying technique origin | Builds on Córdoba/Martínez-Zoroa cascade framework | Builds on Córdoba/Martínez-Zoroa cascade framework |
| Prize intent | OpenAI says it will not claim the $1 million award | Not reported as seeking the Millennium Prize |
Market Impact: What This Means for OpenAI’s Roadmap
For OpenAI, the timing is not an accident. The announcement arrives weeks after the company shipped GPT-6 Astra, and it explicitly positions the unreleased model used for the proof as a step beyond that release. Framing a math result this way lets OpenAI make a capability claim that is, in theory, independently checkable through the Lean formalization, which is a stronger form of evidence than a benchmark score the company reports on its own.
That distinction matters commercially. Frontier labs increasingly compete on reasoning benchmarks that are opaque, self-reported, or easy to game, and a mechanically verifiable proof of a famous open problem is a harder claim to dismiss than a leaderboard number. If the Navier-Stokes proof holds up under review, it becomes a reference point OpenAI can point to in enterprise and research sales conversations about agentic reasoning capability, well beyond what any chat benchmark could carry. If it does not hold up, or if the credit dispute with Buckmaster and Alpöge deepens, the episode instead becomes a cautionary tale about rushing capability announcements ahead of academic norms around priority and peer review.
Competitive Landscape: Anthropic, Google DeepMind, and the Math Race
The involvement of an Anthropic researcher, Levent Alpöge, on the rival Euler result is itself notable. It shows frontier labs are not the only players producing AI-assisted advances on hard formal math, and that individual researchers moving between industry and academia are becoming a meaningful channel for this kind of work, independent of which company employs them. Google DeepMind’s own history with formalized math, including its documented findings on AI agents gaming proof scoring systems, has made it a reference point other labs are measured against when they claim a formally verified result.
None of the major labs have publicly disputed that AI systems are now capable of producing serious, if imperfect, progress on Millennium-tier problems. The disagreement in this case is narrower: who gets credit, whether OpenAI’s system independently reached its result, and whether the release cycle around the announcement gave the mathematical community enough time to check the work before it became a headline. That is a governance and norms problem as much as a technical one, and it will likely shape how the next lab announces its own formal-math result.
The AI-Slop Problem in Formal Proofs
Buckmaster’s characterization of an early LLM-generated proof draft as difficult to read and in need of substantial cleanup echoes a broader pattern researchers have flagged across AI-assisted mathematics this year: models can generate proof-shaped text quickly, but turning that text into something a human referee or a Lean checker will accept often still requires significant expert intervention. That gap between “the model produced something” and “the model produced a correct, checkable proof” is exactly why the Lean formalization, not the press release, will be the thing that ultimately settles whether OpenAI’s Navier-Stokes claim stands.
What Happens Next: 5 Predictions
- Independent mathematicians and Lean specialists will spend weeks to months checking OpenAI’s formalization line by line before the community reaches consensus on whether the proof is correct.
- Expect a formal writeup or commentary from Charles Fefferman or another Clay Institute-affiliated mathematician addressing whether OpenAI’s result matches the official problem statement precisely enough to count.
- OpenAI and Buckmaster/Alpöge are likely to publish a joint or parallel clarification of the timeline to settle the credit question, given how much scrutiny the priority dispute has already drawn.
- Rival labs, including Google DeepMind and Anthropic, will likely respond with their own formal-math demonstrations in the coming months, using the same playbook of a Lean-checkable result rather than a self-reported benchmark.
- Regardless of how the verification process ends, expect this episode to become a case study in how AI labs should (or should not) announce unverified formal results, given the criticism over OpenAI holding a press call before independent mathematicians could review the paper.
Frequently Asked Questions
Did OpenAI actually solve the Navier-Stokes Millennium Prize Problem?
OpenAI claims its internal model produced a proof showing that smooth 3D fluid flow can develop a singularity in finite time, which would resolve the problem. The claim has not yet been independently verified by the mathematical community, and outlets including TheNextWeb and Quanta Magazine have reported that outside mathematicians have not fully reviewed the underlying proof.
Will OpenAI collect the $1 million Millennium Prize?
No. OpenAI has said it does not intend to claim the prize money from the Clay Mathematics Institute.
What is the difference between the Navier-Stokes and Euler equations in this story?
Navier-Stokes describes fluid flow with viscosity, or internal friction. Euler equations describe the same kind of flow without viscosity. OpenAI’s announcement addressed Navier-Stokes, while Buckmaster and Alpöge’s separate result, published about 12 hours earlier, addressed a forced version of the Euler equations.
Who are Tristan Buckmaster and Levent Alpöge?
Tristan Buckmaster is a mathematician at NYU. Levent Alpöge is a researcher at Anthropic. The two published a Lean-verified proof on a forced 3D Euler equations problem shortly before OpenAI’s Navier-Stokes announcement, which led to a public dispute over credit and timing.
What model did OpenAI use to produce the proof?
OpenAI has described it only as an internal, unreleased model that is significantly more capable than GPT-6 Astra. The company has not given the model a public name or release date.
Has any Millennium Prize Problem ever actually been solved?
Yes. The Poincaré Conjecture was solved by Grigori Perelman in the early 2000s. He declined both the $1 million prize and the Fields Medal he was offered for the work. It remains the only Millennium Prize Problem formally resolved and accepted by the mathematical community.
Why does a Lean formalization matter for verifying the claim?
Lean is a formal proof language that lets a computer mechanically check whether each logical step in a proof is valid. A Lean-checked proof is harder to fake than a written argument alone, but mathematicians still need to confirm that the formalized statement in Lean accurately captures the real-world claim being made, which is why independent review is still ongoing.
Is this the same team that built GPT-6 Astra?
OpenAI has said the proof was produced by an internal system built on a next-generation model beyond GPT-6 Astra, without detailing whether it was the same research team. Mark Chen, OpenAI’s chief research officer, was the executive who addressed questions about the effort on a press call.
