OpenAI Says It Solved a Millennium Prize Math Problem — But Verification and a Data Dispute Loom

OpenAI announced on September 8, 2026 that it has produced a solution to the Navier-Stokes Millennium Prize Problem, one of seven unsolved mathematics problems designated by the Clay Mathematics Institute, each carrying a $1 million reward. The Navier-Stokes equations describe how liquids and gases flow, and the associated existence-and-smoothness problem asks whether solutions to these equations always behave well or can sometimes blow up in finite time. The question has stood open since the Millennium Prize list was created (The Verge; Clay Mathematics Institute).

In a blog post titled "On the Navier–Stokes Millennium Prize Problem" published at openai.com/index/navier-stokes-solution/, OpenAI said the solution was produced using an internal AI model more powerful than its newly released GPT-6 Astra, running alongside 10,000 concurrent agents — essentially, thousands of independent instances of the model working on different parts of the problem in parallel (The Verge). The company said training of this internal model began on August 28, and that the model showed unprecedented performance across its benchmarks, including in mathematics. The post also states that the solution includes a formal proof written in the Lean proof assistant, a tool that mechanically checks each logical step in a mathematical argument to ensure no gaps exist (OpenAI).

The claim has not yet been independently verified. As of reporting on September 8, OpenAI had not made the underlying proof publicly available or submitted it to peer review, and mathematicians raised questions about the claim (Firstpost). A Lean-formalized proof, if eventually released, would carry a different evidentiary weight than an unformalized argument, since Lean's kernel mechanically checks each inference step. But the formalization itself has not been shared.

The announcement is complicated by a dispute over data access. On September 7, one day before OpenAI's announcement, NYU mathematics professor Tristan Buckmaster published findings on a problem related to Navier-Stokes, co-authored with Levent Alpöge, a researcher at Anthropic (The Verge). After learning that OpenAI was aware of his and Alpöge's progress, Buckmaster contacted the company and raised concerns about whether OpenAI had accessed their Codex data — records of their interactions with OpenAI's coding tool — in pursuit of its own Navier-Stokes solution.

OpenAI's response has not fully closed the matter. In its September 8 announcement, the company said no specific user data was accessed to solve the problem, but added it could not rule out that de-identified data from users' product usage helped improve its models (The Verge; OpenAI on X). OpenAI staff member Sebastien Bubeck said the company did not see Buckmaster and Alpöge's work until it was released publicly, and that the two proofs differ significantly. Buckmaster countered that OpenAI was openly admitting it used training data from a period after his and Alpöge's result was found (The Verge).

The distinction OpenAI draws between "specific user data" and "de-identified data from users' product usage" is doing substantial work in this exchange. If Buckmaster or Alpöge interacted with Codex or other OpenAI products in the course of their research, the question of whether aggregated or de-identified traces of that activity flowed into model training becomes difficult to answer definitively. OpenAI's phrasing acknowledges this gap rather than closing it. The fact that Alpöge is a researcher at Anthropic, a competitor in the foundation-model space, adds an additional layer of sensitivity to the data-access question, though neither party has framed it in those terms.

The timing is notable on its own terms. Buckmaster and Alpöge's results appeared on September 7. OpenAI announced its solution on September 8, having begun training the relevant model on August 28. Whether the model's training incorporated information derived from Buckmaster and Alpöge's pre-publication work is the crux of the dispute, and the available statements from both sides leave the question open.

OpenAI said it does not plan to claim the $1 million prize associated with solving the Navier-Stokes problem (The Verge). The Clay Mathematics Institute's award process requires publication in a refereed mathematics journal of international repute and a waiting period of at least two years, followed by consideration by the institute's Scientific Advisory Committee. OpenAI has not indicated whether it intends to submit the proof for that process.

Several things would need to happen for this claim to move from announcement to accepted result. The proof, including its Lean formalization, would need to be released for community inspection. Independent mathematicians would need to verify both the mathematical arguments and the correctness of the Lean formalization. And the data-access dispute between OpenAI and Buckmaster would need to be resolved, or at least clarified to the point where the provenance of key ideas in the solution is traceable. None of these steps has been completed.

In my view, the most consequential element of this story may not be the mathematics itself but the infrastructure that produced it. An internal model exceeding GPT-6 Astra in capability, orchestrated across 10,000 concurrent agents and trained for roughly eleven days before yielding a result on a Millennium Prize problem, describes a type of automated mathematical reasoning system that, if the claim holds, moves AI-assisted theorem proving into a qualitatively different regime. The verification lag, the data-provenance dispute, and the absence of a shared artifact for community scrutiny are the immediate friction points. How quickly each is resolved will determine whether this announcement becomes a milestone or a footnote.