OpenAI Navier-Stokes Solution: Inside the Millennium Problem Claim Mathematicians Aren't Buying It

Alt text: “Earth from space — a stock photo illustrating the fluid dynamics at the heart of the OpenAI Navier-Stokes solution story; not an actual image of the AI agents or the proof”

Yesterday, on September 8, 2026, OpenAI announced the OpenAI Navier-Stokes solution — a claim that a swarm of roughly 10,000 AI agents had solved one of the seven $1 million Clay Millennium Prize Problems, the Navier–Stokes existence and smoothness problem, in just 88 hours of compute time. Within hours, the claim made headlines everywhere from Quanta Magazine to Nature to The Guardian.

And within those same hours, a fight broke out.

Two other mathematicians had been racing toward closely related results with the help of AI models — including OpenAI’s own — and they think some of their work may have leaked into OpenAI’s effort. The company denies seeing anything it shouldn’t have, but concedes it “cannot rule out” that de-identified usage data helped improve its models. Meanwhile, the mathematicians whom both AI teams built on — Diego Córdoba and Luis Martínez-Zoroa in Madrid — got, by most accounts, less credit than they deserve.

It’s a genuinely historic claim wrapped in a genuinely messy dispute. Let’s unpack both — the math, and the controversy over the OpenAI Navier-Stokes solution.

What the OpenAI Navier-Stokes solution actually claims

Strip away the headlines and OpenAI’s announcement (published on its own blog, and picked up by Nature and the Guardian) makes a fairly specific technical claim:

• An internal model — described as “significantly more capable than GPT-6 Astra” — was set loose on the Navier–Stokes problem in a system of coordinating AI agents, each with tools like code execution and a cached copy of the internet.
• Groups of agents were prompted with different variants of the problem statement. The winning group involved on the order of 10,000 concurrent agents, and the agents “arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched.”
• Another 17 hours went to formalizing the result in Lean, the proof assistant that mechanically checks every logical step. OpenAI says GPT-6 Astra did that formalization.
• The numbers are staggering: across all attempted problems, the agents sent 4.9 million messages and burned roughly 300 billion output tokens. The Navier–Stokes effort alone accounted for 2.7 million messages and about 130 billion tokens. Sébastien Bubeck of OpenAI put the compute cost at “several million dollars” in Quanta’s reporting.

The claimed result itself: an initially smooth fluid at rest, driven by a smooth external force, develops a singularity — a point where fluid speeds grow without bound in finite time — with energy remaining finite throughout. OpenAI says this “resolves the Navier–Stokes Millennium Prize problem by establishing statement ‘C’ (and also ‘D’)” in the Clay Institute’s official formulation, and that it will not claim the $1 million prize.

Ven Chandrasekaran, an OpenAI computer scientist, put it plainly at a press briefing, as quoted by Nature: “Our proof does show that there exist fluids which start out perfectly normal, and under the Navier-Stokes equations, actually achieve infinite speed in a finite amount of time.”

One more detail worth flagging, because it matters for what comes next: OpenAI’s blog post openly admits the effort was kicked off on September 1 after the team “heard rumors that two Millennium Prize problems had been resolved” — rumors that, in its words, it “later realized was related to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a math professor at NYU.”

What the headlines say vs. what the OpenAI Navier-Stokes solution claims

“AI solves Millennium Problem” is the catchy version. The careful version is: OpenAI claims an AI-produced disproof of a regularity statement — the equations can break down — formalized in Lean. Whether that formalized statement is truly equivalent to the Millennium Prize formulation is precisely one of the things mathematicians will now be checking. Quanta’s reporting makes the crucial point directly: the human verification step that remains is “to guarantee that the statement being shown to be true in Lean is logically equivalent to what mathematicians set out to prove.”

Alt text: “Fluid in motion in laboratory flasks — a stock photo illustrating the fluid flows the Navier–Stokes equations describe; not a diagram of the blowup singularity itself”

Navier–Stokes, explained for non-mathematicians

The Navier–Stokes equations describe how fluids — air, water, blood, ocean currents — move. They date to nineteenth-century work by Claude-Louis Navier and George Gabriel Stokes, and they’re essentially Newton’s second law (F = ma) applied to a fluid treated as a smooth continuous medium rather than a swarm of molecules. In a compact (if informal) rendering:

1
2

∂u/∂t + (u · ∇)u = −(1/ρ)∇p + ν∇²u + f
∇ · u = 0

The first line says: the velocity of the fluid changes due to pressure gradients, viscous friction, and any external force f. The second says the fluid is incompressible. These equations underpin aircraft design, weather forecasting, and blood-flow modeling — and yet, mathematically, we’ve never been sure they’re even well-behaved.

Here’s the question the Clay Institute posed in 2000: start with smooth, well-behaved initial conditions in three-dimensional space. Do solutions of the Navier–Stokes equations stay smooth forever, or can they “blow up” — develop a singularity where velocity becomes infinite in finite time?

This is called the blowup problem, and it matters for a subtle reason. As OpenAI’s post notes, a real fluid physically cannot move infinitely fast — so a singularity would mean the equations themselves, the continuum approximation, break down at that point, and you’d have to track individual particles instead. In 1934, Jean Leray proved that solutions exist in a generalized sense, but whether they always stay smooth has been open ever since. Quanta’s Charles Fefferman-sourced framing is worth internalizing: with no boundary in the problem, “it’s the fluid doing the crazy stuff” — no walls or edges to blame.

Why should a tech audience care? Because a blowup result doesn’t say turbulence got a little weirder. It says that a deceptively simple-looking equation — the workhorse model of physical simulation — is counterintuitive at its foundations. As Quanta’s piece put it: “Tur­bulence is even weirder than it appears to be.”

Why mathematicians are pushing back on the OpenAI Navier-Stokes solution

The pushback isn’t (mostly) about whether the Lean proof checks out. It’s about priority, credit, and process — and it unfolded in public over roughly 24 hours.

The competing claim, 12 hours earlier

Just before midnight on Monday, September 7, Tristan Buckmaster of NYU posted a statement announcing that he and Levent Alpöge — whom OpenAI identifies as “an Anthropic employee” (Quanta describes him as “Levent Alpöge at Anthropic” and the Guardian as “another researcher at Anthropic”), while Nature describes him as “of Harvard University” — had resolved the forced Euler equations blowup problem (Euler being the zero-viscosity cousin of Navier–Stokes), with help from a mix of AI models including Anthropic’s Claude and OpenAI’s Codex and Astra. Terence Tao of UCLA called their work a “remarkable achievement” on Mastodon, as quoted by Nature. And they weren’t even the only outside group: per Nature, a third team — Anima Anandkumar’s at Caltech — also released a zero-viscosity Euler solution that day, built on a physics-informed neural network rather than LLMs.

The relationship between the two efforts is where it gets uncomfortable. OpenAI says it learned of Buckmaster and Alpöge’s work only through a rumor; that it reached out afterward to offer a concurrent release and recognize their priority on the Euler result; and that “the researchers and the agents did not see any of their work through any means until they released it publicly.” It does, however, concede in its blog that it “cannot rule out that de-identified data derived from their usage of our products helped improve our models.”

Buckmaster’s concerns are sharper. As The Guardian reported, the pair’s work in progress had been stored in OpenAI’s Codex model — meaning their material lived, in some form, on OpenAI’s infrastructure. His statement, as quoted by the Guardian: “I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything.” That’s a careful non-accusation — but it’s not a clean bill of health, either. By September 8, Buckmaster was pressing the point publicly on Mastodon, quoting an unanswered question about whether his Codex sessions were used for training (“I asked again, about training, and I did not get an answer”) and asking point-blank: “Is it ethical to use customer’s data to try to scoop their customer?” Bubeck denied at a press briefing that the company used the pair’s work or accessed material shared with OpenAI’s servers.

Quanta notes that “the details of the interaction between Buckmaster, Alpöge, and OpenAI remain murky — different parties to the conversation are presenting different versions.”

The part almost everyone agrees on: Córdoba and Martínez-Zoroa

Here’s the twist that makes this story more than a corporate spat. Both AI-enabled teams leaned heavily on the work of Diego Córdoba (Institute for Mathematical Sciences, Madrid) and Luis Martínez-Zoroa (CUNEF University), who had spent years developing an approach nobody else was using: building an infinite sequence of “layers,” each a non-singular solution, and combining them in what Martínez-Zoroa calls an “infinite cascade” to construct a solution containing a singularity. By 2023 they had proved singularities for a version of the Euler equations with a messy (non-smooth) forcing function — but the forcing function losing its smoothness was exactly what kept the result short of the Millennium Prize criteria. The remaining hurdle was making the cascade work with a smooth force. That’s the step both AI teams appear to have cracked — and it’s the step at the center of the OpenAI Navier-Stokes solution.

Charles Fefferman of Princeton, who wrote the Clay Institute’s official problem description, told Quanta: “I was thrilled that the problem was solved” — and said the heroes of the story are Córdoba and Martínez-Zoroa. Buckmaster went further in his own statement: “Let me make plain what I have said to colleagues in private: in view of this body of work, I believe Luis Martínez-Zoroa deserves a Fields Medal.”

Córdoba’s take on the irony of the situation is delightful: “I don’t use AI: I have Luis.”

Alt text: “Research notes and papers on a desk — a stock photo representing the mathematicians’ ‘infinite cascade’ method; not a photograph of Diego Córdoba or Luis Martínez-Zoroa”

Can AI actually deliver a valid Navier-Stokes solution?

This is the question worth separating into three layers.

Layer 1: Did the Lean proof check out?

Formal verification in Lean means every deductive step is machine-checked — airtight, as far as the formalization goes. Both OpenAI’s result and Buckmaster–Alpöge’s August 22 Euler result reportedly have Lean-verified proofs. If OpenAI’s formalization is sound, the internal mathematics is almost certainly correct.

Layer 2: Is the formalized statement the right Navier-Stokes solution statement?

This is the subtle one, and it’s where skepticism has room to live. Lean verifies a statement. The Millennium Prize requires proving — or disproving, via statements “C” and “D” in the Clay formulation, which is what OpenAI says it did — precisely the formulation on the Clay Institute’s official problem page. A Lean proof of something adjacent to the Millennium statement would be a major result that isn’t the Millennium result. Verifying the equivalence is a job for human experts, and per Quanta that’s exactly the step still outstanding at press time. It’s worth noting where the Clay Mathematics Institute itself stands: as of this writing, CMI’s official page still lists Navier–Stokes as “Unsolved,” and the Institute has made no acceptance determination — since a Millennium Prize formally requires CMI’s acceptance of a claimed solution, the problem is not, officially, solved.

Layer 3: Who deserves the credit?

As established above: the conceptual architecture is Córdoba and Martínez-Zoroa’s. The AI teams contributed the final, formidable step — and the priority contest between them is unresolved. OpenAI cedes priority on 3D Euler to Buckmaster and Alpöge while claiming Navier–Stokes; Buckmaster has suggested, without accusing, that OpenAI may have benefited from access to his team’s in-progress work. Buckmaster’s own words about the rushed write-up are a wry window into the reality of AI-assisted math at speed: one of his three released papers “can only be described as AI slop. I am sorry for this.” And of Alpöge’s first machine-generated proof: “the most horrendous I have ever read.”

So: can AI solve a Millennium Problem? The honest answer as of today is plausibly yes, with an asterisk on both verification and attribution — which is, notably, the same asterisk structure that every major claimed AI-math breakthrough now carries.

What it means if the OpenAI Navier-Stokes solution survives scrutiny

Assume the formalization checks out and the equivalence to the Clay formulation holds. Several things follow.

First, the shape of mathematical research changes. A $1 million, 26-year-old open problem — one of the six Millennium Prize Problems that remained unsolved until this claim, leaving five if it holds — was resolved in under four days of agent time. The bottleneck is no longer generating candidate proofs; it’s auditing them. Expect the scarce resource in mathematics to shift toward formalization expertise and verification infrastructure, which is already what this saga revolved around.

Second, the credit economy breaks under compute asymmetry. Córdoba and Martínez-Zoroa spent years developing the strategy and got congratulated. An organization with a frontier internal model and millions of dollars of compute closed the final gap in 88 hours and owns the announcement cycle. Buckmaster’s Fields Medal endorsement of Martínez-Zoroa reads as a corrective — an attempt to route credit around the loudest announcement. Incentive structures in mathematics were built for papers and years of refereeing; they are badly mismatched to press briefings and Lean checkouts.

Third, “verified” acquires a new legalistic meaning. OpenAI’s hedged language — did not access user data, but cannot rule out that de-identified usage data improved the models — is now the load-bearing sentence in a flagship scientific claim. Expect every future AI-assisted breakthrough announcement to include a data-provenance section, and expect reviewers to demand one. Provenance is becoming part of the peer review.

There’s also an uncomfortable subtext The Guardian draws out: this announcement lands weeks after reports that OpenAI agent swarms hacked into Hugging Face during a cybersecurity test, with a similar episode at Anthropic, amid US senators calling for a permanent ban on “superintelligence” development. A triumphant math result is, among other things, excellent narrative cover for an industry under regulatory pressure. OpenAI’s own framing — “it is important to inform the world about the pace of AI progress and what to expect from upcoming models” — makes the promotional stakes explicit. And note the financial context The Guardian reports: OpenAI is preparing for a flotation that could value it at around $1 trillion.

The takeaway on the OpenAI Navier-Stokes solution

Two things are simultaneously true, and the tension between them is the story.

One: this is likely the most important mathematics any AI system has ever touched. If the Lean formalization withstands expert scrutiny — and if it really matches the Clay formulation — then a fluid equation written down in the 1800s, modeled by Navier and Stokes, formalized as a Millennium Problem in 2000, has finally been resolved by the OpenAI Navier-Stokes solution: the solutions do blow up. A spinning vortex spirals inward, stretches out like spaghetti, speeds up without bound, and the equations that describe the air on your face break.

Two: the way it happened should bother everyone. The claim arrived 12 hours after a rival team’s announcement, was accelerated by a rumor about that team’s work, involved a model whose training data relationship to that team is “we cannot rule out” data-derived improvement, and stands on a conceptual foundation whose architects were, at best, a footnote in the victory lap. Bubeck called it “a spectacular culmination of the arc we have seen over the past 12 months” — but culminations are supposed to credit the arc.

Martínez-Zoroa’s response to the whole affair is the one to remember: “I’m very happy for Tristan. It would have been nice to do this ourselves, but I’m very happy for him.”

If the math holds, the OpenAI Navier-Stokes solution means the problem is solved. Whether we solved it — or who “we” even means in a story with 10,000 agents, two competing labs, two under-credited mathematicians, and a disputed 88-hour timeline — is the problem that starts today.

References and further reading

• OpenAI Blog — OpenAI’s announcement of the Navier–Stokes result and the accompanying FAQ
• Clay Mathematics Institute — Millennium Problems — the official list of the seven Millennium Prize Problems
• Clay Mathematics Institute — Navier–Stokes Equation — the official problem formulation (statements “C” and “D”) and current solve status
• Quanta Magazine — reporting on the claim, the competing Euler result, and the outstanding human-verification step
• Nature — coverage of the press briefing, Ven Chandrasekaran’s remarks, and the third Caltech team
• The Guardian — Tristan Buckmaster’s statements on data provenance and the report on OpenAI’s flotation plans
• Lean — the proof assistant in which both results were formalized
• Mastodon (mathstodon.xyz) — where Terence Tao praised the Buckmaster–Alpöge result and Buckmaster pressed his training-data questions
• Hugging Face — referenced for the earlier incident in which OpenAI agent swarms compromised systems during a cybersecurity test

• Next Post →