OpenAI says a system involving roughly 10,000 concurrent AI agents has produced a solution to the OpenAI Navier-Stokes existence and smoothness problem, one of mathematics’ seven Millennium Prize Problems. The company has released a written proof and a Lean formalisation, while the Clay Mathematics Institute says the problem has ‘apparently been settled’ but stresses that its formal evaluation process is deliberately unhurried.
OpenAI Navier-Stokes work claims a finite-time singularity
The Navier-Stokes equations describe the motion of fluids such as water and air. The Millennium problem asks, in one accepted formulation, whether sufficiently smooth three-dimensional solutions always remain smooth or whether a singularity can develop in finite time. OpenAI says its proof establishes the latter possibility under the relevant smooth-forcing formulation.
The company published the result on 8 September 2026 together with a formalised proof in Lean, a proof assistant that checks whether each logical step follows from precisely stated assumptions. OpenAI says the mathematical construction balances several effects so that the external force remains smooth while the fluid velocity becomes unbounded.
That is a very different claim from a numerical simulation showing turbulent behaviour. A Millennium Prize solution has to establish a mathematical theorem. The Lean formalisation provides a machine-checkable representation of the argument, but the significance and correctness of the overall construction still have to be examined by mathematicians and through the Clay process.
The system used an internal model beyond GPT-6 Astra
OpenAI says the discovery effort used an internal model that was significantly more capable than GPT-6 Astra and was still being trained when the project began. The company launched coordinated agent groups after hearing rumours on 1 September that two Millennium problems might have been resolved and decided to test the system across the remaining open problems.
The group that OpenAI credits with the Navier-Stokes resolution involved on the order of 10,000 concurrent agents. Those agents could read a cached version of the internet, run code, communicate within groups and explore different approaches. OpenAI says it later cross-pollinated useful ideas between groups, with Codex consolidating promising intermediate results into follow-up prompts.
According to the company, the agents reached the Navier-Stokes result about 88 hours after the first agents were launched. Lean formalisation and verification took another 17 hours using GPT-6 Astra. For the Navier-Stokes effort alone, OpenAI reports 2.7 million agent messages and roughly 130 billion output tokens.
Ten thousand agents did not vote on one answer
The scale can sound like 10,000 copies of a chatbot independently solving the same equation and taking a majority vote. OpenAI describes something more like a distributed research programme. Different groups pursued different routes, shared intermediate ideas and were redirected as useful results emerged.
That distinction is important because more agents do not automatically produce better science. LiveAIWire recently covered a benchmark in which AI agents tried to improve training algorithms 290 times and made them worse in 124 runs. Large-scale search creates many candidates, including failures. Value comes from the mechanisms that identify, test and consolidate the rare useful paths.
Mathematics is unusually suitable for that kind of verification because a proof can, in principle, be reduced to formal logical steps. A correct formalisation cannot make a false theorem true merely by sounding convincing. The difficult part is ensuring that the formal theorem actually corresponds to the Millennium problem and that all assumptions match the intended mathematical statement.
Clay says the problem has apparently been settled
On 10 September, the Clay Mathematics Institute issued an unusually positive statement. It said it was contemplating an announcement that the Navier-Stokes problem had ‘apparently been settled’ and welcomed the prospect of new understanding as the work is analysed and interrogated.
That language is significant because Clay created and administers the Millennium Prize Problems. It is still not the same as awarding the US$1 million prize or formally declaring the process complete. Clay’s own rules are designed to prevent a dramatic announcement from becoming an instant prize decision.
Under those Millennium Prize rules, a proposed solution normally has to appear in a qualifying publication, at least two years must pass, and the solution must achieve general acceptance in the global mathematics community before Clay considers detailed prize evaluation. The institute does not accept direct submissions of proposed solutions.
Clay’s current Millennium Problems page also continues to list Navier-Stokes among the active problems. That status is consistent with an important proposed solution being under scrutiny rather than a completed prize award.
OpenAI itself says it is not claiming the prize
OpenAI has explicitly said it does not intend to claim the Millennium Prize. That does not reduce the mathematical importance of the work if it survives review, but it helps separate three questions that can easily collapse into one headline: whether the proof is correct, whether the Millennium problem is considered resolved by the community and whether Clay awards the prize.
The first question can be examined immediately by mathematicians and proof assistants. The second takes time because researchers need to understand the argument, try to break it and determine whether it addresses the intended problem without hidden gaps. The third follows Clay’s rules and institutional process.
This is why the most accurate current formulation is that OpenAI says it has solved the problem and Clay says the announcement appears to have settled it, while formal recognition remains pending. A week of excitement cannot substitute for the years of scrutiny that the prize rules deliberately require.
Formal proof changes the verification problem, not the need for humans
Lean can verify that a formal proof is logically valid relative to its definitions and foundational system. It cannot by itself decide whether the formal statement is the right translation of every informal requirement, whether a published exposition explains the key ideas well or how the result fits into decades of mathematical literature.
Human mathematicians therefore remain central. They need to read the construction, compare it with existing results, check the correspondence between the formal and conventional statements and understand which ideas are genuinely new. Formal verification can make some categories of hidden algebraic or logical error much harder to sustain, but it is part of the review chain rather than a replacement for the community.
LiveAIWire’s report on an AI system that found hundreds of errors in published machine-learning papers illustrated the same principle from the other direction. Automated checking can direct attention and catch objective faults, while human experts remain responsible for deciding what those faults mean for the scientific claim.
AI had already produced a more modest mathematics surprise
The Navier-Stokes announcement arrives only weeks after another high-profile AI mathematics result. LiveAIWire reported that an Anthropic research system failed to solve the Riemann Hypothesis but produced a new 67.2% lower-bound result on a related question. That episode was notable precisely because the failed grand objective still yielded a narrower result mathematicians considered valuable.
The OpenAI claim is on a different scale. If the Navier-Stokes proof is accepted, it would resolve the stated Millennium problem itself rather than produce progress around its edges. It also provides a test case for whether massive multi-agent research can generate work that withstands the most demanding kind of external scrutiny.
Success would not imply that the same approach can solve every open scientific problem. Mathematics offers unusually crisp verification. In biology, materials science or climate research, an agent can propose a theory or experiment but nature still has to supply the final answer. The cost of checking may be much higher than running another proof assistant.
The compute bill is part of the scientific story
OpenAI says all attempted high-impact problems in the project generated 4.9 million messages and around 300 billion output tokens. That scale makes the result as much a story about research infrastructure as about one model. The system coordinated enormous parallel search, communication and formal checking over several days.
The company has not framed those figures as a general recipe for mathematical discovery. A resource-intensive method that succeeds on a famous problem might be justified as a research demonstration while remaining impractical for ordinary academic work. Future systems will need to show whether the same quality can be obtained with less compute or whether scale is a central ingredient.
There is also a selection effect. A public success tells us which project worked, while failed branches are less visible. OpenAI’s disclosure of message and token counts helps reveal the size of the search space, but independent analysis will be needed to understand which parts of the process contributed most to the final proof.
The next phase is interrogation, not celebration alone
Mathematics advances through adversarial reading. Researchers will try to simplify the argument, find unstated assumptions, reproduce the Lean environment, test edge cases and relate the construction to known partial results. If the proof is sound, that scrutiny will strengthen it. If a gap exists, the same process should expose it.
The Clay statement captures that posture unusually well: excitement paired with deliberate delay. A result can be extraordinary and still need years before institutional recognition. That is not bureaucratic foot-dragging. It is how a prize intended to recognise achievements of historic magnitude protects the meaning of the award.
For AI research, the stakes are broader. A verified solution would be powerful evidence that agentic systems can contribute original frontier mathematics rather than merely remix known results. It would also raise new questions about credit, publication and how human researchers should interact with systems capable of exploring many reasoning paths in parallel.
For now, the defensible conclusion is strong enough without exaggeration. OpenAI has released a proof and formalisation that it says resolve Navier-Stokes; Clay has publicly described the problem as apparently settled; and the global mathematics community has begun the slow work of deciding whether the claim survives. The AI sprint lasted days. The verification process is supposed to take much longer.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.