OpenAI says it cracked Navier-Stokes in 88 hours of agents

On September 8, 2026, OpenAI published an analytical proof and a Lean formalization claiming that an initially smooth fluid can develop a finite-time singularity under the Navier-Stokes equations, after about 88 hours of work by some 10,000 agents, amid a priority dispute with human mathematicians.

On September 8, 2026, OpenAI announced that it had produced, via an internal agent system, a solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems posed in 2000 by the Clay Mathematics Institute. According to the company's research publication, the proof shows that a three-dimensional incompressible fluid, starting from a smooth state at rest, can develop a finite-time singularity under a smooth external force, while retaining finite energy. The OpenAI Navier-Stokes announcement places, for the first time according to Nature, a genuinely major open problem in mathematics under the banner of a resolution claimed by computer.

The sequence is brutal in its duration. OpenAI says it launched the effort on September 1, after rumors that Millennium Problems had been solved elsewhere. The agents are said to have succeeded on September 5, about 88 hours after the start, after which GPT-6 Astra reportedly formalized and verified the proof in Lean over an additional 17 hours. Mark Chen, head of research, called the result a "significant milestone," while putting the computation cost at "millions of dollars." Sam Altman said the lab wanted to know if "one of ours" could also get there, after hearing about progress linked to Anthropic.

What OpenAI concretely says about OpenAI Navier-Stokes

The Navier-Stokes equations, inherited from the work of Claude-Louis Navier and George Gabriel Stokes in the 19th century, describe fluid motion as a continuous medium. They are used in aircraft design, weather forecasting, and the study of blood flow. The open question is whether this continuous approximation can blow up mathematically. In other words, whether velocity can grow without bound in finite time, despite viscosity tending to smooth out motion. Such a singularity would mark a limit of the model, since a real fluid cannot reach infinite velocity.

OpenAI claims to have established statements "C" and "D" of the Clay Institute's official formulation. The described solution is a vortex that curls inward and stretches out "like spaghetti," with a central region that shrinks while accelerating. The technical difficulty, the company writes, is that the blow-up must arise from the fluid's own motion, not from an infinite force introduced by hand. The terms for acceleration, pressure, momentum transfer, and viscosity must become large while balancing each other precisely, so that an external force remains smooth even as velocity diverges.

Two nuances matter for readers looking into OpenAI Navier-Stokes. First, the published proof concerns a regime with smooth external force (forced version), not the "bare" equations without forcing. Numerama notes that whether a blow-up obtained with an external force "counts" for the Millennium Prize is part of the debate. Second, OpenAI explicitly writes that it does not intend to claim the prize of one million dollars. Martin Bridson, president of the Clay Mathematics Institute, welcomed "an exciting day," while noting that the evaluation is "deliberately unhurried," with journal publication, community acceptance, and then a wait of at least two years.

How 10,000 agents produced the proof

The model used, still in training according to OpenAI, is said to be "significantly more capable" than GPT-6 Astra, launched less than a week earlier. The effort mobilized a system of coordinated agents, with access to a cached version of the web and code execution, split into groups of varying sizes. The group that found the Navier-Stokes resolution numbered on the order of 10,000 competing agents, under the same monitoring and isolation safeguards as other frontier evaluations.

Before reaching full scale, about 1,000 agents are said to have first solved, in 50 hours, an "easier" version linked to the Euler equations (Navier-Stokes without viscosity), in the unforced case. This success convinced the lab to concentrate resources on Navier-Stokes. The agents cross-checked their approaches, including via Codex to consolidate useful ideas. On the Navier-Stokes problem alone, OpenAI reports about 2.7 million messages and 130 billion output tokens. Across all problems combined, the total is said to reach 4.9 million messages and nearly 300 billion tokens.

Ven Chandrasekaran summarized, at the press briefing reported by Nature, the physical stakes. The proof would show that there exist fluids that "start out perfectly normal" yet reach "infinite velocity in finite time" under Navier-Stokes. Because a real fluid cannot behave this way, this suggests that, in certain regimes, the equations cease to be a reliable mirror of physical reality. Sébastien Bubeck spoke of a "spectacular culmination" of twelve months of accelerating model performance on problems of increasing complexity.

The priority dispute surrounding OpenAI Navier-Stokes

A few hours before the announcement, mathematicians Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) made public their own results, obtained with the help of Claude, Codex, and Astra. They describe in particular a blow-up for the Euler equations in a forced setting, and say they also have progress toward the more general case. Anima Anandkumar (Caltech) and her collaborators also published, the same day, a solution to the zero-viscosity problem via a physics-informed neural network, rather than via a general-purpose large language model.

Buckmaster published a lengthy statement questioning OpenAI's timing. He fears the lab may have picked up an unorthodox approach after learning of their method, and questions the possible use of data linked to Codex. Bubeck rejected "false and inflammatory" accusations, published screenshots of exchanges, and OpenAI wrote that it "did not use their prompts or results to steer our models or agents." In its public note, however, the company acknowledges that it "cannot rule out" that de-identified data derived from usage of its products may have helped improve its models, while asserting that the proofs differ and that the Euler result is not the same (forced in Buckmaster–Alpöge's case, unforced on OpenAI's agents' side before moving to forced Navier-Stokes).

Sam Altman recounted that the rumors of an Anthropic resolution had been experienced as a slap in the face, and that the lab had then proposed a coordinated publication and even suggested Buckmaster as lead author of a rewrite, a proposal that did not go through. Buckmaster, for his part, highlighted the ethical question on Mastodon. Is it acceptable to use training data that postdates a client's discovery in an attempt to surpass them? Terence Tao, Fields Medalist, praised Buckmaster and Alpöge's work as "remarkable," while regretting, according to AFP, that the search for mathematical solutions is being entrusted entirely to models rather than to human–AI collaboration.

What OpenAI Navier-Stokes changes for science and the labs

For readers looking into OpenAI Navier-Stokes, the essentials boil down to this. What: OpenAI publishes an analytical proof and a Lean formalization of a finite-time singularity for Navier-Stokes (statements C and D, forced regime), without claiming the Clay Prize. Who: agents of an internal model more capable than Astra, Mark Chen and Sébastien Bubeck on OpenAI's side, Buckmaster and Alpöge in parallel, the Clay Institute for future validation. When: effort launched September 1, agent resolution on the 5th, Lean on the 6th, public announcement on September 8, 2026. What it changes: the debate is no longer only about models' ability to solve olympiad problems, but about scientific priority, community verification, and the pace at which a lab can convert compute into "millennium" results.

The Lean formalization attests to a machine-verified logical consistency, Scientific American notes via Numerama, but does not equate to acceptance by the mathematical community. The Clay Institute requires peer-reviewed publication, general acceptance, and then a two-year wait. Until then, the OpenAI Navier-Stokes case will remain both a demonstration of agent power and a test of scientific governance. Who gets credit for what, how one publishes, and what happens when two teams — one human aided by AI, the other almost entirely agentic — hit the same historic wall days apart.

Sources

• OpenAI — On the Navier–Stokes Millennium Prize Problem, September 8, 2026
• Nature — OpenAI claims huge maths breakthrough on a famed 'Millennium Problem', September 8, 2026
• France 24 / AFP — OpenAI says it solved a major math problem in less than four days, September 9, 2026
• franceinfo — OpenAI says it solved the Navier-Stokes equation, September 9, 2026
• Numerama — OpenAI claims to have solved part of a millennium math problem, amid plagiarism accusations, September 8, 2026
• Numerama — A mathematician accuses OpenAI of trying to appropriate his breakthrough, September 8, 2026

Frequently asked questions

What did OpenAI announce on September 8, 2026 about Navier-Stokes?

The company published an analytical proof and a Lean formalization claiming that an initially smooth fluid at rest can develop a finite-time singularity under the Navier-Stokes equations, with a smooth external force and finite energy. It presents this as resolving statements C and D of the Clay formulation, without claiming the Millennium Prize.

What does OpenAI Navier-Stokes concretely mean?

It is the claim, dated September 8, 2026, that an OpenAI agent system produced a machine-assisted proof of blow-up (singularity) for the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The proof still needs to be read and accepted by the mathematical community.

How many agents and how much time were needed?

OpenAI states about 10,000 competing agents, a resolution in about 88 hours starting September 1, followed by 17 hours of Lean formalization via GPT-6 Astra. For Navier-Stokes alone, it cites about 2.7 million messages and 130 billion output tokens, for a computing cost of "millions of dollars."

Will OpenAI receive the Clay Institute's one million dollars?

No, according to its own publication. It writes that it does not intend to claim the prize. The Clay Institute requires peer-reviewed publication, general acceptance by the community, and then a wait of at least two years after that publication.

Why is there talk of plagiarism or priority?

Tristan Buckmaster and Levent Alpöge published, the same day, results on related equations (notably forced Euler) and questioned OpenAI's timing and the possible reuse of an approach or data. OpenAI denies having directed its agents with their prompts or results, acknowledges it cannot rule out an effect of de-identified data on model improvement, and states that the proofs and even the precise results differ.

Does the proof concern the equations without external force?

The OpenAI proof described for Navier-Stokes falls within a regime with smooth external force (statements C and D). The agents had previously handled an unforced version for Euler (zero viscosity). The forced/unforced distinction is at the heart of the technical debate and the comparison with Buckmaster–Alpöge's work.