On a press call earlier this month, OpenAI announced that one of its internal models had completed a proof related to one of mathematics’ most difficult open questions. At the time, no proof was shown publicly. That has now changed, and the published record substantially narrows what OpenAI is claiming.

The company has posted a detailed write-up, a PDF version of the paper, and a formal proof encoded in the Lean theorem prover. The Lean files are available in a public repository. Since Lean proofs can be mechanically checked by anyone with the right tools, outsiders do not need to take OpenAI’s word for the mathematical argument. OpenAI says the formalisation and verification step consumed another 17 hours of computation using GPT-6 Astra.

A striking singularity claim

The result concerns the Navier-Stokes equations, the equations used to describe the motion of fluids. What OpenAI says it found is a fluid that starts at rest in a smooth state and then develops a singularity in finite time. A singularity, in this context, means that an important physical quantity, such as velocity or pressure, becomes unbounded, so the equations can no longer describe the flow in the ordinary sense.

In OpenAI’s framing, that result establishes statements “C” and “D” in the official formulation of the Navier-Stokes Millennium Prize problem. Two paragraphs after that claim, the company adds: “We do not intend to claim the Millennium Prize for this result.” The decision is not a footnote. For any researcher or organisation that truly believed it had settled a Clay Mathematics Institute problem, a $1 million prize and a permanent place in the history of mathematics would be powerful incentives to collect. Passing on the prize is the clearest possible clue about how OpenAI rates its own work against the formal prize conditions.

Why the prize may not apply

The Millennium Prize problems were set by the Clay Mathematics Institute in 2000, and each carries a $1 million award. Only the Poincaré conjecture has been solved so far. The Navier-Stokes problem is regarded as one of the hardest among them, because it asks for a rigorous mathematical understanding of behaviour that can only be approximated numerically in most real settings.

The key issue is not just the proof but the setting. OpenAI says its fluid has “a smooth force applied to it” throughout, from the initial rest state until the moment the singularity appears. External forces are not just a technical detail. In much of the mathematical literature on Navier-Stokes, versions of the question differ sharply depending on whether a force is present and, crucially, whether that force is smooth and well behaved.

OpenAI itself draws this distinction. It mentions separate work by Levent Alpöge and Tristan Buckmaster on the Euler equations, a related but inviscid fluid model. According to OpenAI, the variant Alpöge and Buckmaster addressed was “the unforced version, where no external force is applied.” That phrase marks the distance between that work and OpenAI’s own forced-fluid setup.

The open mathematical question, therefore, is whether a singularity formed under an applied force satisfies the conditions in the Clay Institute’s statement of the Navier-Stokes problem. That is not a matter of computation; it is a question to be answered by mathematicians. It is also a question that may take some time to adjudicate, since the formal prize criteria are written in a way that is much more cautious than a press release.

How the proof was generated

Behind the announcement lies an unusual computational effort. OpenAI says that about 10,000 concurrent agents worked on the problem over roughly 88 hours, from 1 to 5 September. The effort consumed 2.7 million messages and around 130 billion output tokens. Those numbers describe a research instrument that is unlike a single chatbot session or a large reasoning model.

The model that did the heavy lifting was not GPT-6 Astra, which OpenAI shipped a few days before the announcement. It was an internal model that OpenAI says is significantly more capable. GPT-6 Astra’s role was limited to verifying the result after the fact. That verification, according to OpenAI, required an additional 17 hours and produced the Lean formalisation that outsiders can now run themselves.

The total compute bill is not itemised in the published materials, but OpenAI has previously described its large internal efforts in terms of millions of dollars. The Navier-Stokes proof may therefore be one of the most expensive mathematical arguments ever constructed, at least in terms of compute alone.

A change in research culture

There is a broader message in the paper’s provenance. The proof does not come from a single, enormous chain-of-thought session. It comes from a swarm of models working in parallel for three and a half days, with the full force of a large infrastructure spend aimed at one question. That is a different kind of instrument from the ones that most researchers and even most well-funded laboratories can deploy.

The boundary of useful AI research appears to be moving, at least partly, from questions of raw model intelligence to questions of scale. How many agents can be harnessed, how long they can run, and how much compute one is willing to spend on a single problem may matter as much as the architecture of the model itself. This shift is not necessarily good news for mathematics as a discipline. If the only way to prove certain types of results is to spend millions of compute dollars, then the mathematical enterprise gains a strange new dependence on infrastructure budgets.

It also makes independent reproduction harder. A Lean proof can be checked at low cost, because the check is a deterministic local operation. But generating a proof of this complexity in the first place is not something an individual mathematician with a laptop can repeat. In classic mathematical practice, a proof is persuasive because any sufficiently trained person can read and verify it. With a machine-generated proof, that social mechanism is replaced by a tool. In this case the tool is available and open, but the process that created it is not.

The credit question

The published page also includes a striking concession on credit. Earlier reports noted that Buckmaster had asked whether OpenAI pursued directions derived from his unpublished work with Alpöge. At that time, OpenAI denied the implication. The newly published page, however, credits both Alpöge and Buckmaster for concurrent work on the forced Euler problem. It even offers to recognise their priority in a joint announcement.

That is an unusual move for any group that thought the allegation was baseless. A company that had complete confidence in its own originality would not typically offer a public priority agreement within days of the question being raised. The sequence of events suggests that legal or reputational pressure may have been part of the conversation. OpenAI has not, in the published text, admitted that it used their work. But the careful language invites readers to draw their own conclusion.

What can now be verified

Separating what is verifiable from what is not has become easier with this release. The mathematics can be checked by anyone who is willing to run the Lean proof and follow the machine-generated argument. That part is open and reproducible. Whether the result satisfies the precise conditions of the Millennium Prize is not determined by publication alone. The Clay Institute and the mathematical community will need to consider whether the forced setting fits their formulation.

Provenance, meanwhile, remains outside the reach of any formal verification. A question about whether a model absorbed unpublished ideas during training is not something that Lean can adjudicate. It is a claim about a training process that is hidden inside OpenAI. No amount of mathematical formalism can make that provenance transparent, because the relevant data is not public and the training pipeline is not inspectable.

The net result is a strange mixture of openness and opacity: a publicly checkable

Source: TNW | Openai News