OpenAI Says It Solved a Famous Fluid Dynamics Problem — But a Credit Fight Is Already Brewing

OpenAI says it has used AI to produce a solution to the Navier-Stokes equation, one of mathematics’ most famous unsolved problems and a Clay Millennium Prize challenge worth $1 million. The claim matters because it would mark a major step for AI in advanced research, but it is already being overshadowed by a dispute over whether OpenAI raced ahead after learning that outside mathematicians were close to similar progress.

The company announced the result on Monday, saying the achievement came after more than 1,000 AI agents worked on the problem for over 50 hours and consumed millions of dollars in computing resources. Yet the announcement immediately drew pushback from mathematician Tristan Buckmaster, who says OpenAI knew about his own work, moved aggressively to beat him to the punch and even suggested arrangements around credit that he believes were inappropriate.

At stake is not just who deserves recognition for a breakthrough on a century-old fluid dynamics puzzle. The episode also highlights a new and potentially messy reality: as AI systems become more capable of generating proofs, the field may face more disputes over authorship, timing and what counts as original discovery.

What OpenAI says it accomplished

OpenAI says its system found a mathematically formalized solution to a long-standing problem in fluid mechanics known as the Navier-Stokes equation. The equation is used to describe the motion of liquids and gases, including the behavior of water, air and turbulence, and it has resisted a proof in the general case for nearly two centuries.

The company framed the result as a landmark demonstration that AI can contribute to cutting-edge mathematics rather than simply assist with routine calculations. According to OpenAI researchers, the proof was checked in Lean, a formal language increasingly used to verify mathematical arguments with machine precision.

Sebastien Bubeck, a mathematician and OpenAI researcher, said in a briefing that the company began training a new model with stronger mathematical abilities on August 28. He said the team turned more attention to Navier-Stokes after hearing rumors that Anthropic researchers were making progress on a related direction.

Bubeck said the team committed substantial additional resources to the problem and eventually found a Lean-formalized proof after a long run involving thousands of model-driven attempts. He described his reaction as disbelief at first, followed by confirmation that the proof held together.

OpenAI also emphasized how expensive the work was. Mark Chen, the company’s head of research, said the compute bill ran into the millions of dollars, underscoring how far the company was willing to go to push AI into harder mathematical territory.

Why the Navier-Stokes equation matters

The Navier-Stokes equation is one of the most famous open problems in pure and applied mathematics because it sits at the intersection of abstraction and everyday physics. Engineers use fluid dynamics every day, but proving that the equations behave nicely in all circumstances remains one of the deepest unsolved questions in the field.

It is also one of the Clay Millennium Prize Problems, a list of seven famously difficult mathematical challenges selected by the Clay Mathematics Institute. Each carries a $1 million prize for a correct solution, though the real value is the prestige attached to solving one of the discipline’s defining mysteries.

If OpenAI’s claim withstands scrutiny, it would be more than a publicity moment. It would suggest that AI agents can do something closer to frontier mathematical research: explore ideas, test variants, formalize arguments and persist across a long, computationally expensive search in a way that resembles a large research team.

How OpenAI says the proof came together

OpenAI says the breakthrough emerged from a large-scale agentic search, with many models working in parallel on related mathematical ideas before converging on a proof that could be checked formally. The company described the process as computationally intense and unusually expensive compared with prior mathematics projects.

What is Lean, and why does it matter?

Lean is a programming language and proof assistant used to express mathematics in a way a computer can verify step by step. In this context, a Lean-formalized proof matters because it reduces the chance that a claimed discovery is merely a persuasive argument with hidden gaps. It is a way of turning abstract reasoning into something machine-checkable.

That distinction is important in advanced mathematics, where a proof can be elegant, controversial or incomplete depending on how carefully it is written and validated. Formal systems like Lean are increasingly seen as a bridge between human mathematical intuition and computer verification.

Why this is different from ordinary AI assistance

OpenAI’s framing suggests this was not a chatbot casually answering a hard problem, but a highly coordinated research workflow. The system reportedly explored many approaches across a large number of agents, using the scale and parallelism of modern AI infrastructure to attack a question that human mathematicians have wrestled with for generations.

• Hundreds or thousands of candidate paths can be explored simultaneously.
• Promising ideas can be formalized faster than a human-only team might manage.
• Failed branches can be discarded without exhausting a small research group.
• Compute, not just raw model quality, becomes a decisive factor.

Why the credit dispute escalated so quickly

The controversy began because another mathematician says OpenAI’s timing was not coincidental. Tristan Buckmaster, a mathematician at New York University, and Levent Alpöge, a researcher at Anthropic, posted materials on Monday claiming a significant advance in an area relevant to the Navier-Stokes problem. The pair said their work used several AI tools, including Anthropic’s Claude and OpenAI’s Codex.

Buckmaster then posted a statement alleging that OpenAI had become aware of his and Alpöge’s progress last week and soon began directing substantial resources toward the same problem. He also said he asked whether OpenAI had accessed the pair’s Codex logs, and that the company told him it had not looked up user data.

According to Buckmaster, the broader exchange did not end there. He claimed OpenAI offered proposals about how the result might be announced, including one in which he would publish a paper saying the Navier-Stokes problem had been solved by an internal OpenAI model, but without including Alpöge’s name. Those claims, if accurate, would add an unusually sensitive authorship dispute to an already high-stakes scientific announcement.

Buckmaster’s account suggests he believed OpenAI was trying to shape both the timing and the attribution of a result that he and Alpöge believed was connected to their own work.

OpenAI, for its part, rejected the suggestion that it used the pair’s prompts or proof materials. Bubeck said the researchers and agents involved in the project did not see the outside work before it was published publicly. He also stressed that OpenAI recognized the prior work done by Buckmaster and Alpöge on a related problem.

What OpenAI and the outside researchers are disputing

The disagreement is not only about who got to a result first. It is also about whether one side learned enough from the other’s private or semi-private work to influence its own research direction.

That distinction matters because AI research can generate a wide range of logs, prompts and intermediate outputs. If those materials are shared across teams, accessed directly or indirectly, or simply inferred from published hints, the line between inspiration and appropriation can become blurry.

The competing claims, side by side

The following table summarizes the main points of contention as presented in public statements.

How much did compute matter?

OpenAI’s own description suggests the answer is: a great deal. The company said the Navier-Stokes effort required far more compute than it has spent on previous mathematical problems, signaling that brute-force scale and persistent search were essential parts of the process.

That is notable because AI progress is often discussed as a matter of model architecture or data quality alone. Here, the story appears to be about concentrated research infrastructure: many agents, long runtimes, formal verification and the willingness to spend millions in pursuit of one result.

This may also hint at a broader shift in AI labs. If future mathematical breakthroughs depend on expensive agentic searches, then access to enormous compute budgets could matter as much as raw mathematical talent. That would make elite AI research even more concentrated in the hands of a few well-funded companies.

What happens next?

The most immediate next step is independent verification. In mathematics, especially on a problem this famous, a company’s internal announcement is only the beginning. Other experts will want to inspect the proof, reproduce the formalization and determine whether the claim holds up under public scrutiny.

There is also the matter of authorship and attribution. If Buckmaster and Alpöge’s related work is judged to have contributed meaningfully to the breakthrough, the community will likely debate how much credit they deserve and how OpenAI should describe its own role. Those questions may not have a neat answer, especially in a field where AI systems and human researchers are increasingly working in overlapping workflows.

Possible outcomes from here

1. The proof is independently validated and OpenAI gets broad recognition for the technical result.
2. Mathematicians conclude the proof is incomplete, weakening the claim.
3. The work is accepted, but the credit controversy continues over who influenced what.
4. The episode becomes a precedent for how AI-assisted mathematics should be published and attributed.

Why this story could shape AI research norms

Even if the underlying proof survives every technical challenge, the surrounding dispute could be the more lasting story. AI-assisted research is moving into areas where publication timing, private prompts, internal logs and model access may be nearly as important as the final theorem.

That creates familiar scientific problems in a new form. Traditional research already deals with competition, priority claims and disputes over authorship. But AI introduces additional complications: automated search can accelerate work so quickly that two groups can converge on the same idea within days, and the tools themselves may leave detailed records that are hard to interpret fairly.

For now, OpenAI is presenting the event as evidence that AI can push into the hardest corners of mathematical research. Its critics, meanwhile, are warning that the race to claim those achievements may be moving faster than the norms meant to govern them.

As more of mathematics becomes machine-assisted, the question may not only be whether AI can prove something new. It may also be who gets to say they found it first, and what kind of evidence will be required to settle that dispute.

Update note: This is a developing story and may be revised as more details emerge or as the proof is independently evaluated.