OpenAI has announced what could become one of the biggest demonstrations of artificial intelligence in mathematics: its internal AI system has produced a proposed solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems.
The company says its system demonstrated that an initially smooth three-dimensional fluid can develop a singularity in finite time, resolving the problem by establishing one of the official problem’s disproof cases. OpenAI released both a written proof and a formalized version in Lean.
But there is a major caveat.
The mathematical community has not yet formally accepted the result as a completed Millennium Prize solution.
And the announcement has become controversial almost immediately, because mathematicians Tristan Buckmaster of NYU and Levent Alpöge of Anthropic were simultaneously working on related fluid equations using AI tools. Buckmaster has raised questions about how OpenAI arrived at its result so quickly and about the handling of credit and data. OpenAI denies accessing their specific work.
The result is therefore not simply an AI breakthrough story.
It is becoming a test of how scientific discovery should work when AI systems can conduct mathematical research at enormous speed.
OpenAI Claims a Solution to a 90-Year-Old Problem
The Navier–Stokes equations describe how fluids move.
They are fundamental to fluid dynamics and have applications ranging from aircraft design and weather modeling to blood-flow research.
But one basic mathematical question has remained unresolved for roughly 90 years: if a fluid starts with smooth, well-behaved conditions, do the equations always produce a smooth solution, or can the mathematics develop a singularity in finite time?
The Clay Mathematics Institute selected the question as one of its seven Millennium Prize Problems in 2000.
Each problem carries a $1 million prize for an accepted solution.
OpenAI says its system found a finite-time singularity in a three-dimensional incompressible fluid with a smooth external force. In the company’s description, a vortex spirals inward and becomes increasingly stretched, with the fluid velocity eventually becoming unbounded while the total energy remains finite.
If independently verified and accepted under the Clay Institute’s rules, that would represent an extraordinary mathematical result.
But that process has not happened yet.
OpenAI Used Thousands of AI Agents
The scale of the computation is one of the most striking parts of the announcement.
OpenAI says it launched the project on September 1 after hearing rumors that researchers were making progress on Millennium Prize problems.
The company then created a multi-agent research system in which different groups of AI agents attacked different versions of the mathematical problem.
For the Navier–Stokes effort, OpenAI says roughly 10,000 concurrent agents were involved.
The agents could communicate within groups, run code and access a cached version of the internet. OpenAI says the groups were deliberately given different formulations and approaches so they could explore the problem independently.
The agents first tackled a related Euler-equation problem.
After that work produced a promising result, OpenAI redirected additional computational resources toward Navier–Stokes.
The company says the final resolution was produced on September 5, approximately 88 hours after the initial agents were launched.
A further 17 hours were used to formalize and verify the result in Lean using GPT-6 Astra.
Across its broader mathematical experiment, OpenAI says the agents sent 4.9 million messages and generated roughly 300 billion output tokens. The Navier–Stokes effort alone accounted for about 2.7 million messages and approximately 130 billion output tokens.
That is an enormous amount of computation devoted to one mathematical question.
The AI That Found the Result Was Not GPT-6 Astra
There is another important detail that could easily be missed.
OpenAI says the system that generated the proposed Navier–Stokes solution was an internal model significantly more capable than GPT-6 Astra.
GPT-6 Astra was used later to help formalize and verify the proof in Lean.
This means the breakthrough should not simply be described as “GPT-6 Astra solved Navier–Stokes.”
Instead, OpenAI’s description is closer to:
A new internal AI research system generated the solution, while GPT-6 Astra helped formalize and verify it.
That distinction is important because OpenAI says the internal model was still being trained as the experiment took place and that its performance was continuing to improve.
The company has not publicly identified the model in the same way it has identified Astra.
Why Mathematicians Are Questioning the Announcement
The controversy centers partly on timing.
Before OpenAI announced its result, NYU mathematician Tristan Buckmaster and Levent Alpöge had released research involving related fluid equations.
Their work focused on the Euler equations, which are closely related to Navier–Stokes but omit viscosity.
They used AI tools including Anthropic’s Claude as well as OpenAI’s Codex and Astra.
Nature reported that Buckmaster and Alpöge had also been working toward the more general Navier–Stokes problem.
OpenAI says it began its own effort after hearing rumors that two researchers had solved versions of Millennium Prize problems.
The company says it did not see Buckmaster and Alpöge’s work before it became public and that the proof it produced is significantly different.
However, the researchers’ proximity to the OpenAI project has created a dispute over priority, attribution and whether information from AI-assisted research could have influenced the company’s models indirectly.
OpenAI Denies Using Their Specific Research
OpenAI directly addressed the controversy in its publication.
The company says its researchers and agents did not see Buckmaster and Alpöge’s work through any means before the researchers published it.
OpenAI also says no specific user data was accessed to solve the problem.
However, the company included an important qualification: it said it could not completely rule out the possibility that de-identified data derived from product usage may have indirectly helped improve its models.
That distinction is at the center of the wider debate.
There is a difference between:
“OpenAI accessed this research and gave it to the agents.”
and:
“AI models may have benefited indirectly from information contained in previous product usage.”
OpenAI says the first did not happen.
The second, it says, cannot be completely ruled out.
That uncertainty is particularly sensitive when the AI company itself is competing with researchers who use its products to conduct mathematical research.
The Credit Question Is Even More Complicated
The dispute is not only about data.
It is also about who gets credit when AI and humans collaborate on mathematical discovery.
Buckmaster has publicly criticized OpenAI’s handling of the situation, while OpenAI says it attempted to recognize his and Alpöge’s work and offered a concurrent release.
OpenAI says that after completing its proof and Lean verification, it contacted the researchers because it believed the earlier rumor concerned their work.
The company says it offered them visibility into the prompts and proof and recognized their priority for the forced Euler problem.
Reports from the researchers describe the interactions differently and have raised concerns about how academic credit should work when a large AI company enters a research area already being explored by academics.
The disagreement highlights a problem that traditional academic publishing was not designed for.
A human researcher usually has a fairly clear chain of notebooks, drafts, conversations and publications showing how an idea developed.
With AI systems, that chain becomes much more complicated.
Thousands of agents may explore thousands of ideas simultaneously.
Determining who contributed the key insight can become extremely difficult.
Solving the Problem Is Not the Same as Winning the $1 Million Prize
Another important distinction is the difference between OpenAI’s claim and an officially recognized Millennium Prize solution.
OpenAI says it has solved the problem.
But the Clay Mathematics Institute has its own requirements for recognizing a solution.
The proof must survive expert scrutiny and satisfy the institute’s rules before the prize can be awarded.
OpenAI explicitly says it does not intend to claim the $1 million Millennium Prize for its result.
That means the current story should be described carefully.
It is accurate to say:
OpenAI says its AI system solved the Navier–Stokes Millennium Prize Problem.
It is too early to say:
The Millennium Prize has officially been awarded to OpenAI.
Those are very different claims.
Why the Navier–Stokes Result Matters Beyond Mathematics
Even if the proof ultimately requires revisions, independent verification or additional work, the AI achievement could still be significant.
The important development is not simply the answer to one mathematical problem.
It is the research process.
OpenAI used thousands of agents to explore a difficult problem simultaneously, combine useful intermediate discoveries and then use another AI system to formalize the resulting proof.
That resembles a miniature automated research organization.
Instead of one model answering one prompt, OpenAI created a hierarchy of computational researchers.
Some agents explored.
Others communicated.
Codex consolidated useful ideas.
A stronger internal model produced the key result.
GPT-6 Astra then helped formalize the proof.
This is much closer to an AI research laboratory than a conventional chatbot.
The Bigger Problem: Can Humans Verify AI-Generated Mathematics?
The Navier–Stokes announcement also connects directly to the growing debate around AI-generated mathematics.
A proof can be enormous.
An AI system may produce a logically valid argument that contains hundreds of pages of technical reasoning, formal code and intermediate lemmas.
Human mathematicians still need to determine whether the proof is correct and understand what it actually means.
Formal verification helps with logical correctness.
But it does not necessarily answer every question about mathematical significance, elegance or originality.
That is why the Navier–Stokes controversy is arriving at the same time as a broader debate about “proof abundance.”
AI could potentially produce mathematical proofs faster than humans can absorb them.
And once AI systems begin competing with researchers on open problems, questions about attribution and originality become unavoidable.
AI May Be Changing the Rules of Mathematical Discovery
The most important lesson from the OpenAI announcement may therefore have little to do with the $1 million prize.
It is about the changing economics and speed of research.
OpenAI says its agents used approximately 130 billion output tokens on the Navier–Stokes effort alone.
Thousands of AI researchers worked in parallel.
A result emerged in days rather than decades.
Whether the proof survives expert scrutiny is still a separate question.
But the experiment demonstrates something potentially transformative:
AI systems can now be organized to attack difficult mathematical research problems at a scale that is impossible for a single human researcher.
That changes what researchers may consider feasible.
The Claim Still Needs the Most Important Test: Independent Mathematics
The next stage is not another AI benchmark.
It is independent mathematical scrutiny.
Experts need to examine the proof, verify its assumptions, reproduce the argument and determine whether every step satisfies the precise requirements of the Millennium Prize formulation.
Lean formalization is a powerful part of that process, because a computer can check the formal proof’s logical structure.
But formalization itself must still accurately represent the mathematical claim being made.
And the wider academic community must evaluate the result independently.
Until that happens, the safest description remains OpenAI’s claimed solution, not an officially settled mathematical fact.
OpenAI’s Navier–Stokes Claim Could Be Bigger Than a Single Proof
The extraordinary part of this story is not just that an AI company says it solved one of mathematics’ hardest problems.
It is that the event exposes several questions that AI research will increasingly have to answer.
Who owns an AI-assisted discovery?
How should researchers receive credit?
Can a model’s previous user interactions indirectly influence a later scientific result?
How should AI-generated proofs be verified?
And what happens when thousands of AI agents can explore mathematical ideas faster than human researchers can follow?
OpenAI’s Navier–Stokes announcement may eventually be remembered as a genuine mathematical breakthrough.
But even if the proof is ultimately accepted, the controversy surrounding it will remain important.
Because the future of mathematical research may not simply be about whether AI can solve problems humans could not.
It may be about whether the scientific community can develop trustworthy rules for determining who discovered the solution, how it was discovered, and how humans should understand it.
Frequently Asked Questions
Did OpenAI really solve the Navier–Stokes Millennium Prize Problem?
OpenAI says its internal AI system produced a solution and a Lean formalization showing that a finite-time singularity can occur. However, the result still requires independent mathematical scrutiny and has not been formally recognized as a Clay Mathematics Institute Millennium Prize solution.
What is the Navier–Stokes problem?
It asks whether smooth three-dimensional solutions to the Navier–Stokes equations can remain smooth forever or whether a singularity can develop in finite time under the specified conditions.
How many AI agents did OpenAI use?
OpenAI says roughly 10,000 concurrent AI agents worked on the Navier–Stokes problem.
How long did the AI take?
OpenAI says its agents reached the proposed Navier–Stokes resolution after approximately 88 hours. Lean formalization and verification took another 17 hours.
Did GPT-6 Astra solve Navier–Stokes?
Not exactly. OpenAI says the mathematical solution came from a newer internal model that is significantly more capable than GPT-6 Astra. Astra was used to help formalize and verify the result in Lean.
Why are experts questioning OpenAI’s claim?
The concerns involve independent research by Tristan Buckmaster and Levent Alpöge, the timing of OpenAI’s effort, questions about research priority and attribution, and whether AI systems could have indirectly benefited from information contained in product usage.
Did OpenAI admit using the researchers’ work?
No. OpenAI says its researchers and agents did not see the researchers’ work before it was publicly released. The company did say it could not completely rule out the possibility that de-identified data from product usage had indirectly helped improve its models.
Will OpenAI receive the $1 million Millennium Prize?
OpenAI says it does not intend to claim the Millennium Prize for this result. In any case, a mathematical result would need to satisfy the Clay Mathematics Institute’s requirements before a prize could be awarded.
Why is the breakthrough important for AI?
The project demonstrates a new form of AI-assisted research in which thousands of agents can explore a difficult problem simultaneously, exchange findings, consolidate promising ideas and produce a formally checkable mathematical result.
