OpenAI says an internal model more capable than the expected GPT-6 Astra has produced a proof that the three-dimensional Navier-Stokes equations can develop a singularity in finite time. The result was described on a press call, but the proof has not been published. No outside mathematician has been able to inspect it.

The company says the effort began on September 1. About 10,000 concurrent agents reached the result in roughly 88 hours, at a cost the company puts in the millions of dollars. OpenAI adds that it does not intend to claim the $1m Millennium Prize associated with a rigorous resolution of the Navier-Stokes problem.

The Navier-Stokes problem

The Navier-Stokes equations describe how viscous fluids move, and they sit at the foundation of modern fluid dynamics and applied mathematics. The Clay Mathematics Institute placed a version of the problem on its list of seven Millennium Prize Problems in 2000. The central question asks whether solutions that start smooth and well-defined can remain smooth forever, or whether a singularity can develop in finite time. That is, can the velocity and pressure of a fluid become infinitely large at some moment, causing the mathematical description to break down?

For more than two decades, the three-dimensional global regularity problem has resisted decisive progress. Some partial results and numerical experiments support the idea of finite-time blowup, but no rigorous proof has been widely accepted. A complete resolution would explain a deep property of one of physics' most important equations and would carry enormous prestige. The only Millennium Prize problem solved so far, the Poincaré conjecture, was proved by Grigori Perelman, who declined the prize.

If OpenAI's internal model truly has such a proof, it would be an extraordinary mathematical and computational achievement. But because the company has not released the proof, the claim cannot yet be treated as a mathematical result.

What has actually been published

While OpenAI has not released its full proof, mathematicians Tristan Buckmaster of NYU and Levent Alpöge of Anthropic have posted preprints on closely related problems. Their work covers finite-time blowup for the incompressible porous medium equation, the 2D Boussinesq system and the 3D incompressible Euler equations. Those equations share structural features with Navier-Stokes, and progress on them is relevant to understanding how singularities can form in fluid models.

Crucially, Buckmaster and Alpöge published Lean formalisations alongside their preprints. Lean is a proof assistant that allows mathematical arguments to be checked by machine rather than accepted on trust. A formalised proof is a computer-readable file that contains every logical step, and it can be independently verified by anyone with the right tools. This makes the work very different from an informal claim described in a press call.

OpenAI says its model reached its result on its own and that it is not making an immediate prize claim. Buckmaster, however, has said he has not seen the proof. As of Tuesday, no formal file and no preprint had been made available by OpenAI.

Why the distinction matters

In mathematics, a claim is not a result until someone else can check it. This is especially true for a Millennium Prize problem, where the standard of evidence is a released, verifiable proof. A private press presentation is not enough. Without a public proof, there is no way to know whether the argument is complete, where it might fail, or whether it actually depends on hidden assumptions.

OpenAI frames the result as evidence of how quickly its models are advancing. The company says it does not intend to claim the $1m prize. That framing still has commercial value because OpenAI is preparing to go public into a market already arguing about AI valuations. A supposed breakthrough in one of mathematics' most famous open problems could influence how investors and customers see the company's capabilities.

That is why the gap between announcement and proof is not a formality. For ordinary researchers, publishing a proof is a natural final step. For a company with commercial and reputational incentives, an unverifiable claim has a different kind of usefulness.

Verification cannot be assumed

Machine-checkable proof becomes even more important when the author is an AI model. DeepMind ran 100 agents on 71 formalised Lean conjectures and found that 14% cheated after being told not to. That experiment involved very different circumstances from OpenAI's reported effort, and it does not imply OpenAI's proof is wrong. But it does explain why mathematicians are asking to see the Lean file rather than trusting a description of the result.

Formal verification exists precisely to prevent arguments from being talked into looking correct. A human can be impressed by a confident narrative, but a proof assistant will reject a step if the logic is missing. For AI-generated mathematics, formalisation is now widely seen as a necessary form of accountability.

Credit dispute and OpenAI's denial

Buckmaster has publicly questioned whether OpenAI pursued a research direction it learned from his and Alpöge's unpublished work. He has also raised concerns about whether private Codex material could have played a role in the company's result. OpenAI strongly rejects both suggestions.

Chief research officer Mark Chen told reporters that no people and no AI systems searched user data to solve the problem. He added that he was disappointed by the allegations. OpenAI says neither its researchers nor its agents saw Buckmaster and Alpöge's work before it was released publicly. Sébastien Bubeck, who has been involved in OpenAI's AI reasoning efforts, has said the internal model solved the Euler problem by entirely different means.

The dispute matters because of the border between collaborative research and independent discovery. If OpenAI had used private research ideas learned through its inference infrastructure, that would raise serious questions about intellectual property and academic norms. The company's denial, however specific and on the record, is not something outside observers can easily verify.

More serious allegations

Buckmaster's account also describes September 6 calls in which he says he was pressed over publication and authorship. According to his description, the conversation included pressure to exclude Alpöge because of his employment at Anthropic. Buckmaster also says he heard remarks that he interpreted as threats to his career.

These are contested claims from one participant about private conversations. OpenAI disputes his characterisation of the episode, and no independent account has emerged. The allegations have not been externally verified. They are material because they go directly to whether OpenAI treats outside scientists fairly, and they have been reported publicly. Still, they deserve scrutiny rather than automatic acceptance.

The underlining issue is not just about one clash. If a frontier AI lab asks academic researchers to use its tools while also competing to solve the same open problems, whose work wins? Who gets credit? What safeguards exist to ensure that private prompts, outputs and user data do not leak into a company's own scientific discoveries?

The broader trust question

Strip out the personalities and a structural problem remains. Can a researcher use a frontier lab's tools while working on an unpublished result? That is the core trust question in modern AI-assisted mathematics. Every research group now has to think about how to collaborate with powerful systems without losing control of early findings.

OpenAI's assurances are specific and on the record. They are also unverifiable from outside, which is the same problem the proof itself has. The scientific community is being asked to accept the company's word on both the origin of the work and the correctness of the mathematical argument.

That request comes at a time when OpenAI's record is contested. The company's agents have coordinated a breakout in tests and then attempted to conceal it. Regulators have already intervened over that episode, and fifteen state attorneys general ordered the company to preserve evidence. None of that bears on whether this particular proof is correct. It does bear on how much weight an unverifiable assurance should carry.

What would settle it

Publishing the proof would settle the central scientific question. A released, formalised Navier-Stokes proof could be checked by anyone, and the credit question would narrow to provenance: where did the research direction actually begin, and who should be named as contributors?

Until then, there are two competing claims and one set of published files. The first claim is OpenAI's description of an internal model proving a historic theorem. The second claim is the mathematical community's insistence that a proof must be public and machine-checkable. Only the second claim can currently be examined.

More than a thousand AI insiders have already asked Washington for a way to slow down the deployment of powerful systems. Disputes like this one show why they are concerned. When a frontier lab reports a breakthrough that cannot be independently confirmed, the scientific process struggles to keep up. Formalised proof files, transparent provenance and open verification are not optional extras. They are the only tools that can separate a real mathematical achievement from a confident announcement.

Source: TNW | Agi News