OpenAI’s Navier-Stokes Claim Is Tangled in a Credit Dispute

OpenAI announced on September 8, 2026 that its AI agents had solved the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000. The achievement was almost immediately overshadowed by accusations from NYU mathematician Tristan Buckmaster, who says he and Anthropic employee Levent Alpöge spent nearly a year on the same problem and that OpenAI built on their work without giving them credit. OpenAI denies the claim, but key questions remain unanswered.

What happened

Detail	Fact
Problem solved	Navier-Stokes existence and smoothness (Millennium Prize Problem)
Prize value	$1 million (OpenAI says it will not claim it)
Buckmaster/Alpöge timeline	Nearly one year of AI-assisted work on the problem
Models Buckmaster/Alpöge used	Publicly available models from both OpenAI and Anthropic
OpenAI’s internal model	Unnamed; described as dramatically outperforming the Astra model released last week
Prior Millennium Problems solved	One (before this announcement)

The Navier-Stokes problem asks whether a set of equations describing how fluids like water and air move over time will always produce physically meaningful results, or whether they can, under some conditions, predict something impossible, such as a fluid reaching infinite velocity. On the same day OpenAI made its announcement, Buckmaster posted a proof on Mastodon showing that a simplified version of the equations can in fact break down. OpenAI then presented a proof for the full equations, obtained using an internal model it says far exceeds its recently released Astra model.

Both proofs rely on an approach to the problem pioneered by mathematicians Diego Córdoba and Luis Martínez-Zoroa. According to Brown University mathematics professor Javier Gómez-Serrano, that approach was one of several considered promising, so it is not impossible both teams landed on it independently. But the overlap is striking.

What did Buckmaster actually allege?

Buckmaster posted a document alongside his proof describing his conversations with OpenAI employees after he heard rumors of their work and contacted them. According to him, OpenAI staff offered him two options: he and Alpöge could post their own work and OpenAI would publish its solution the next day, or Buckmaster could collaborate with OpenAI on a paper that left Alpöge off the author list entirely because of his affiliation with Anthropic, OpenAI’s main competitor.

Buckmaster also wrote that he asked OpenAI employees whether its agents had accessed transcripts of his and Alpöge’s sessions with OpenAI models. They denied it. When he asked whether those transcripts had been used in training, he received no answer.

OpenAI Chief Research Officer Mark Chen repeated at a press briefing that no agents or employees accessed those transcripts. But OpenAI’s Sébastien Bubeck acknowledged the team was “inspired to pursue the problem after hearing a rumor” about Buckmaster and Alpöge’s efforts. That admission does not prove data access, but it closes the gap between “completely independent” and “aware of their work.”

Why it matters beyond the math

This episode raises two separate questions that are easy to conflate. The first is whether OpenAI’s agents did anything technically improper by training on or accessing private research transcripts. The second is bigger: what happens to academic mathematics when the resources needed to solve its hardest problems are concentrated at a handful of private companies?

Academic mathematics runs on open collaboration, citation norms, and the ability for any researcher anywhere to build on published work. The episode described by Buckmaster, where he was offered authorship terms that excluded a collaborator for competitive reasons, sits uncomfortably alongside those norms. As MIT Technology Review notes, if OpenAI’s proof did depend on Buckmaster and Alpöge’s direction, then human “research taste” (the judgment to pick a promising approach over others) was still essential. That is a small comfort for human mathematicians, but it does not resolve the structural problem.

For businesses and teams watching the AI frontier, the practical signal is different: AI agents working with internal compute that “dramatically outperforms” publicly available models can now close a year of expert human work in what appears to be a much shorter time. That gap will matter in any field where deep technical research is a competitive moat. We have previously covered OpenAI’s initial Navier-Stokes announcement and the pace of investment pouring into AI coding and reasoning, both of which point in the same direction.

Our take

The math itself is genuinely impressive. A Millennium Prize Problem had stood open for decades, and AI agents appear to have been central to closing it. But OpenAI’s handling of the announcement is a textbook example of how to turn a win into a credibility problem.

Offering a collaborator the choice between “post now and we publish tomorrow” or “join us but without your co-author” is not collaboration. It is pressure. The refusal to answer whether training data included private research sessions is the kind of non-answer that invites exactly the suspicion Buckmaster has raised publicly.

For any business thinking about integrating AI into research or knowledge-intensive workflows, this case is a useful reminder: the more capable your AI tools become, the more important it is to know what data they trained on and what they can access. Opacity at that level is not just an ethics problem. It is a liability.

The structural issue (frontier AI companies controlling the compute needed for the hardest problems) is worth watching closely. Right now it looks like an academic concern. Give it a few years and it could affect any industry where proprietary research is an asset.

What to do about it

1. Audit what your AI tools can access. If agents or models have access to confidential research, internal communications, or client data, document what is in scope and what is not.
2. Ask vendors directly whether your session data is used for training. Silence, as this case shows, is itself an answer worth noting.
3. Follow how major institutions respond to this episode. If academic bodies or journals set new norms around AI-assisted proofs and attribution, those norms will migrate into other fields.
4. If your team does novel research with AI tools, keep a clear record of your prompts, sessions, and results. That record is your evidence of independent work if a dispute ever arises.

The bigger the AI’s role in producing valuable work, the more important your paper trail becomes.

Source: MIT Technology Review