AI solved a Millennium Problem: what OpenAI proved and what it means for business

OpenAI's AI agents produced a proof for one of the seven Millennium Problems, yet the problem is officially still listed as unsolved and nobody has received the prize. On September 8, 2026, the company announced that its internal model and roughly ten thousand agents had constructed a solution of the three-dimensional Navier–Stokes equations that goes to infinity in finite time. The construction comes with an important caveat: the flow is pumped by an external force. The computation took 88 hours of machine time, and translating the proof into formal terms took another 17. The compute cost millions of dollars, against a Clay Mathematics Institute prize of one million.

Below is a breakdown of the story without the hype. What exactly was proved, and under which version of the problem statement. Why machine verification of the proof matters more than the press release. What the compute cost and how that figure compares with the prize. What lies behind the authorship dispute with mathematicians from New York University and Anthropic. And most importantly: which conclusions apply to the everyday tasks of a company considering AI agents of its own. We have a separate breakdown of how the agents themselves are built: AI agents for business.

What exactly happened on September 8, 2026?

In short: ten thousand agents working in parallel assembled a proof of solution blowup in 88 hours, and a machine formally verified every step.

According to OpenAI's publications and the analysis by Simon Willison, the timeline looks like this. The agents were launched on September 1 and produced the result on September 5 — 88 hours of machine computation. Another 17 hours went into translating the proof into formal terms. The outcome is a 166-page manuscript, “Finite Time Blowup for Navier-Stokes,” and a project in the Lean language that checks every logical step automatically.

The scale of the computation is expressed in numbers mathematics is not used to:

A caveat about the 88 hours: that is the time from the launch of the first agents on September 1 to the result on September 5, and the run covered several open problems at once, not only Navier–Stokes.

The model that ran the computation has not been released publicly. Formalization and verification were carried out by GPT-6 Astra — that is, a checking model from the same vendor as the model being checked. All the figures come from OpenAI itself; there is as yet no independent audit of the costs or the logs, and that is worth keeping in mind when reading any headline about a “breakthrough.”

What are the Navier–Stokes equations, and why is a million dollars on offer for them?

In short: they are the equations of fluid and gas motion that engineers have used for a century and a half, even though no mathematical proof of their well-behavedness existed.

The equations describe how water flows through a pipe, how air passes over an aircraft wing, how smoke disperses, and how concrete is mixed. They were derived in the 1820s–1840s; turbine calculations, weather forecasting, and all of computational fluid dynamics rest on them, and engineering practice confirms them every day.

What remained open was the smoothness question: does a solution always exist and stay finite over any interval of time, or does the velocity in an infinitesimally small region become infinite at some moment? Such a moment is called a singularity. The question goes back to Jean Leray's 1934 paper, which is why the problem is sometimes called ninety years old, even though the equations themselves are about two hundred.

In 2000 the Clay Mathematics Institute added Navier–Stokes smoothness to its list of seven Millennium Problems, with a prize of one million dollars for each. So far only one has been solved: the Poincaré conjecture, proved by Grigori Perelman, who declined the award in 2010. For a business reader, what matters is not the sum but the status: this is the class of problem mathematicians spend decades on without getting a result.

Has the problem really been solved?

In short: the proof has been published and machine-verified, but there is no official recognition, and the problem statement comes with an important caveat about the external force.

Clay Institute rules require three things before a claim is even considered: publication in a qualifying venue, two years after publication, and general acceptance of the result by the world mathematical community. Only then does the scientific advisory board convene a special committee of experts. As of September 2026, Navier–Stokes is still listed among the unsolved problems, and the institute's president, Martin Bridson, promised an assessment that would be “deliberately unhurried” and “absolutely rigorous.” OpenAI, for its part, has said it does not intend to claim the million-dollar prize.

The second caveat is subtler and more important. The official statement allows one to prove either that a smooth solution always exists or the opposite — that it breaks down. OpenAI closed the second case in a weakened form: with an external force that pumps energy into the flow. The text of the statement explicitly permits that variant. The community treats the case without an external force as the substantive question, and there is as yet no public evidence that the construction works without one.

The strong side of the announcement is the formal verification. The proof has been translated into the Lean language, where a computer checks every logical transition, and the verification code is published openly. That removes the main risk of a long manuscript: a person can miss an error in a 166-page derivation for months.

What such verification does not provide is confidence that the formalized statement matches the requirements of the official problem statement: smoothness classes, decay conditions, and function spaces are checked by a human, and that work still lies ahead. Nor had the verification been independently reproduced by an outside group at the time of publication.

How much did the proof cost, and why does that figure matter more than the sensation?

In short: the compute cost several million dollars against a one-million-dollar prize, and that is the central economic fact of the whole story.

Estimates diverge, because only OpenAI knows the real cost:

• The company itself speaks of “millions of dollars”: that is how chief research officer Mark Chen described the cost. Sébastien Bubeck added that it is roughly a thousand times more than the company's earlier mathematical experiments.
• Outside estimates based on public pricing. 130 billion output tokens work out to about 6–6.5 million dollars at the average consumer price; with input tokens included, estimates reach 10–40 million. University of Michigan professor Karthik Duraisamy cited a similar figure — about 6 million dollars at retail rates.
• OpenAI's own cost is below retail. In his analysis, University of Michigan professor Karthik Duraisamy estimates the computation at roughly 400 thousand GPU-hours on the company's own fleet — on the order of 1 million dollars, not the market-rate millions.
• The prize is 1 million dollars, and it does not cover the retail price of the compute. Sam Altman responded with a joke on social media: tokens are sold at a loss, and the result “only cost a million.”

A twentyfold spread in the estimates is itself a fact worth remembering: the public price of tokens and the model owner's own cost differ by a multiple. For a company sizing AI up against its own tasks, there is a practical conclusion here. A result of this class is bought with compute at retail prices, and the price rises faster than the value of the answer. In corporate tasks the sums are of a different order, but the pattern is the same: extra millions of tokens rarely compensate for an undocumented process and bad data. We covered how these costs break down in practice in our article on the stages and cost of adopting AI.

What is the authorship dispute about?

In short: mathematicians Tristan Buckmaster and Levent Alpöge had been working toward a similar result for nearly a year and say OpenAI launched its computation within days of word about their work reaching the company.

The timeline looks like this. Buckmaster of New York University and Alpöge, who works at Anthropic, spent nearly a year on the result and in August 2026 obtained blowup proofs for related equations — Euler, Boussinesq, and flow in a porous medium; by August 22 the derivations had been verified in Lean. They announced them publicly on the evening of September 7. By its own account, OpenAI started its computation on September 1, having heard a rumor that competitors were close to a solution, and published its paper on September 8.

The complaints concern two things: the similarity of the approach and the data. Buckmaster and Alpöge used OpenAI and Anthropic models in their work and asked whether their working sessions had ended up in training. Speaking for OpenAI, Sébastien Bubeck replied that the company had not used their prompts or proofs and had not seen their work before publication (Fortune). OpenAI did acknowledge that it cannot rule out one thing: anonymized derived data from use of its products may have improved the models. Separately, Buckmaster described a conversation with company representatives as pressure, and that part remains his word against OpenAI's.

Terence Tao, one of the most authoritative working mathematicians, called Buckmaster and Alpöge's results “a remarkable achievement.” A few days before the announcement he wrote about a risk of a different kind, endorsing a point made by Hugo Duminil-Copin: indiscriminately mining open problems for answers can destroy the environment in which new methods and a new generation of researchers are born.

For business there is an applied part to this story, far removed from mathematics. The question “what happens to what we send into a public model” has stopped being theoretical: it is being asked publicly by people whose profession is precision of wording. What exactly goes out to an external service, what never goes out, and where the data is physically processed — that is decided before work begins, not afterwards. A detailed breakdown is in our article on how to work with neural networks without leaking data.

Does this mean AI will replace experts?

In short: the agents won where a machine checks the correctness of every step, while framing the problem and judging its significance are still done by people.

Consider the mechanics. The problem was formulated by people and had existed in finished form for twenty-six years. The line of attack — constructing a solution that blows up in finite time — was developed by mathematicians in earlier work. The agents' role was to search through an enormous number of variants of the construction, with immediate formal verification of each. Where a human takes a week to check a derivation, Lean returns a verdict in minutes, and brute-force search becomes a viable strategy.

That oracle is exactly what most corporate tasks lack. There is no universal checker for the phrase “suggest how to restructure the sales department,” so the text comes out plausible in form and unverifiable in substance. But an oracle does exist for the tasks a company usually considers boring: reconciling documents against the accounting system, checking a contract against a checklist, monitoring stock levels, processing source documents. There the result can be compared with a reference automatically.

One more detail that gets lost in the headlines: 10,000 agents were needed not because that is a good way to work, but because the problem allows parallel search and the company has its own compute fleet. In a business environment the opposite configuration wins — one or two agents with real access to the systems, real permissions, and an action log. What such an agent is built from is covered in the article how to build an AI agent, and the order of adoption in the piece on rolling out AI agents.

What from this story applies to a company's everyday tasks?

In short: AI pays off where the result can be checked more cheaply than it was produced.

Three conclusions that carry over from mathematics into operational work:

1. Verifiability matters more than model power. The agents won because Lean checked every step of theirs instantly. Before handing a task to a model, answer the question of what you will check the result with: a test, reconciliation against 1C (the accounting and ERP platform most Russian companies run on), a data schema, an acceptance checklist. Without such a check, a working solution is indistinguishable from a nicely written text, however much compute you buy.
2. The economics are counted in full, and at your price list. The same computation costs the model's owner about a million dollars and up to forty at retail rates. You pay retail, and on top of the tokens the budget picks up integrations, verification, and support; that total is what you compare against the cost of doing the work by hand.
3. Confidentiality is settled before the first request. In the authorship dispute, a separate topic was what happens to users' working sessions: OpenAI denies direct use of the data and allows only for the influence of anonymized derivatives. The list of what must not be sent to an external model is written in advance, and it applies equally to mathematicians and to the sales department.

The order of the first steps — pick one task, measure the “before” state, define the boundaries — we covered separately in the article on neural networks for business; for this topic it is the same.

INCUBE AI builds custom systems around a specific process: ERP and CRM, AI agents and integrations, under contract, with data held in Russia and support after handover. News like September's shows the upper limit of what the technology can do, while a company's money comes from the lower one — the work whose result is checked and repeated every day. If you have a task where the result can be compared against a reference, tell us about it: we will map the process and show where AI will pay off and where ordinary automation is enough.

Sources

• Quanta Magazine, September 8, 2026 — analysis of the result, the Lean verification, and the reaction of the mathematical community and Terence Tao
• Simon Willison, “On the Navier–Stokes Millennium Prize Problem” — timeline of the computation, token and message volumes, OpenAI's position and the mathematicians' complaints
• Yahoo Tech, September 8, 2026 — 130 billion output tokens, 2.7 million messages, outside estimates of the compute cost
• Implicator.ai: the Clay Institute's position — the problem is listed as unsolved, variants (C) and (D) with an external force, OpenAI's refusal of the prize
• Millennium Prize Problems rules, Clay Institute — publication in a qualifying venue, a two-year wait, general acceptance by the community
• List of Millennium Problems, Clay Institute — the seven problems of 2000, a 1-million-dollar prize for each
• OpenAI's paper “Finite Time Blowup for Navier–Stokes,” September 8, 2026 — 166 pages, the construction with an external force, reference to alternatives (C) and (D) of Fefferman's statement
• The official problem statement, Charles Fefferman — four solution variants and the conditions on the external force
• Terence Tao's post, September 3, 2026 — a warning about mining open problems, citing Hugo Duminil-Copin
• Karthik Duraisamy, analysis of September 9, 2026 — an estimate of the computation in GPU-hours and a cost of about 1 million dollars
• The Lean repository of the proof — open code of the formal verification
• Fortune, September 8, 2026 — Tristan Buckmaster's complaints, Sébastien Bubeck's reply, Terence Tao's warning