OpenAI has said it found a solution to the Navier-Stokes problem, putting artificial intelligence at the center of one of mathematics’ most daunting unsolved challenges. If the work withstands expert scrutiny and meets the exact standards required for formal recognition, it could represent the second Millennium Prize Problem solved since the prize program began in 2000.

But the headline is not simply “AI solves hard math.” Almost immediately, the claimed result became wrapped in a disagreement over unpublished work, access to user material entered into AI tools, and who should receive credit in any academic publication. The dispute involves Tristan Buckmaster, a mathematician at New York University; Levent Alpöge, an Anthropic researcher who was reportedly pursuing the work in a personal capacity; OpenAI mathematician and AI researcher Sebastien Bubeck; and OpenAI CEO Sam Altman.

The central point remains unsettled: OpenAI has announced a solution, while Buckmaster has raised questions about whether OpenAI’s effort benefited from his and Alpöge’s research activity in OpenAI’s Codex system. Bubeck and Altman have responded publicly with their own accounts, and Bubeck has specifically rejected the allegation that he asked for Alpöge to be removed from a paper. That makes this less a clean victory screen than a high-stakes multiplayer match over provenance, authorship and verification.

Why the Navier-Stokes problem matters

The Navier-Stokes equations are foundational to fluid dynamics. They describe the movement of fluids and gases, which means they matter to questions involving water, air, turbulence and many other physical systems. Yet an equation being useful in science and engineering is not the same as having every mathematical property fully nailed down.

The Millennium Prize version of the problem asks mathematicians to establish whether, in three dimensions, the equations always behave properly for smooth starting conditions or whether they can produce a breakdown known as a singularity. In broad, nontechnical terms, the task is to determine whether the mathematics can reach an impossible or uncontrolled state under conditions where the physical model should remain meaningful.

It is one of seven Millennium Prize Problems established by the Clay Mathematics Institute in 2000. Each carries a $1 million award for a qualifying solution. Before this Navier-Stokes announcement, only one of the seven had been solved: the Poincaré conjecture. That history explains why a claimed solution demands both attention and caution. A result of this scale is not certified by a splashy post, a promising model output, or even a persuasive early manuscript. It must endure close examination by specialists.

OpenAI has said it does not plan to claim the $1 million prize. That does not make the question of credit less important. Formal mathematical authorship, the creation of an auditable proof, and the trail showing how a result came together can carry consequences far beyond a prize purse.

The researchers’ concerns about AI-assisted work

Buckmaster and Alpöge had reportedly been pursuing their own approach to Navier-Stokes and used more than one large language model during that work. The tools named include Anthropic’s Claude, OpenAI’s Codex and OpenAI’s Astra frontier model. Alpöge’s employer is Anthropic, although the work in question was described as personal rather than conducted on the company’s behalf.

Buckmaster has argued that OpenAI appeared to intensify its Navier-Stokes work after learning that a group involving an Anthropic employee was nearing a solution. More seriously, he questioned whether information from the pair’s Codex sessions could have contributed to OpenAI’s own work. He said the researchers had placed drafts from the project into those sessions and asked whether the model had been trained on, or otherwise had access to, that user data.

His account says he was told the model did not retrieve user data, but that he did not receive an answer when he followed up with a question about training. This distinction is crucial. “The model does not look up a user’s current private chat” and “no submitted material can ever influence training or internal development” are different claims. Any controversy involving AI-assisted research can turn on exactly how a product handles data, what settings apply, which services are involved, and what contractual or technical controls govern submitted content.

At present, the public dispute does not itself establish that OpenAI used Buckmaster and Alpöge’s material. It does establish that the researchers are asking for clarity about the route from private AI-tool interactions to a company’s separate research claim. For an industry increasingly eager to place AI systems in coding, science and scholarship, that is a question with broad importance.

The publication and attribution dispute

Buckmaster also described conversations with Bubeck and another unnamed person about how the work might be published and credited. In Buckmaster’s telling, two paths were presented. One would have involved releasing a partial development while OpenAI published its claimed Navier-Stokes solution the next day. The other would have involved Buckmaster authoring a paper that recognized the role of an OpenAI language model while omitting Alpöge from the author list. Buckmaster said he rejected both options.

Bubeck has disputed a key part of that account, specifically denying that he requested Alpöge’s removal from a paper. The disagreement is therefore not merely about a vague sense of who got there first. It concerns a concrete and consequential allegation about authorship.

Mathematics has long had intense norms around priority. Researchers regularly circulate preprints, discuss partial results, and build on ideas from colleagues, but publication records and author lists establish who contributed what. AI complicates that terrain. A language model can suggest transformations, search patterns, summarize known techniques, write code or help researchers test lines of reasoning. None of those functions automatically answer whether the system was a tool, a collaborator-like source of insight, a conduit for confidential material, or some combination of the above.

Those questions become sharper when the AI provider is also attempting to publish work in the same area. It creates the sort of conflict concern that would make any speedrunner pause before accepting a judge’s ruling from their direct rival. Strong documentation, clear data policies and an independent review process are not cosmetic extras in that setting; they are central to trust.

A solution still needs mathematical validation

It is important to separate a company’s announcement from final recognition by the mathematical community. OpenAI’s claim is momentous, but the Navier-Stokes problem has a very demanding formal target. A proof must precisely address the stated conditions of the Millennium problem, withstand attempts to find gaps, and be evaluated over time by qualified mathematicians.

That process is not a flaw in the system or an attempt to diminish AI. It is how mathematics works. Proofs are meant to be independently checked. In fact, the use of AI makes rigorous transparency even more valuable. Reviewers need to know what the actual proof is, which steps are machine-generated or machine-assisted, whether every inference can be reproduced without proprietary model access, and whether the presentation omits assumptions hidden inside code, prompts or computational workflows.

The episode also arrives during a wider push to use AI for frontier research. Companies want models that do more than generate fluent explanations: they want systems that help produce verifiable advances in mathematics, materials science, biology and physics. The same push is part of why questions about access, competition and accountability matter. Elsewhere in the technology sector, platform power and self-preferencing concerns have already become regulatory flashpoints, as seen in Google’s planned EU search changes following a Digital Markets Act penalty. Research tools raise their own distinct issues, but the common thread is whether a dominant platform can set rules that users cannot meaningfully inspect.

What needs to happen next

There are several separate questions that should not be mashed into one oversized “did AI solve it?” button prompt:

  • Proof: Does OpenAI’s submission fully solve the Navier-Stokes Millennium Prize Problem under its published criteria?
  • Reproducibility: Can independent experts inspect and verify the argument without relying on inaccessible proprietary systems?
  • Provenance: What records can establish how OpenAI’s result developed and whether outside unpublished work played any role?
  • Data handling: What protections applied to the Codex sessions used by Buckmaster and Alpöge, including any rules around training and internal access?
  • Attribution: If related research independently or jointly contributed, how should that contribution appear in papers and public accounts?

There may be answers that satisfy some of those questions and not others. A valid proof could still sit alongside a legitimate disagreement over credit. Conversely, a clear data-policy explanation would not by itself prove that a mathematical manuscript is complete. Treating each issue separately will matter if the field wants a resolution that is more durable than a social-media exchange.

For now, OpenAI’s announcement is a remarkable claim with an equally remarkable burden of proof. The Navier-Stokes equations have resisted generations of mathematicians, and any genuine resolution deserves meticulous evaluation. The accompanying dispute is a reminder that in AI-era science, the route to a discovery can be nearly as consequential as the discovery itself.