OpenAI has presented a 166-page potential solution to the Navier-Stokes problem, but mathematicians cited in a Scientific American report question whether it follows the approach the company attributed to Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. The debate also points to a second uncertainty: what role the large number of AI agents used to develop the work actually played.
A disputed link to earlier mathematics
OpenAI said its potential proof followed the approach of Córdoba and Martínez-Zoroa. Several groups of experts now disagree with that characterization. Gonzalo Cao-Labora, a mathematician at EPFL, says the similarities are only surface-level and that the work takes a different approach when examined in detail.
That distinction matters because a claim about following an existing approach is more specific than saying two efforts look alike. If the detailed methods differ, the connection OpenAI described may not capture how the new work was developed. The source account does not settle which interpretation is right; it reports a disagreement among mathematicians.
The potential proof is written in Lean, a formal verification language. Lean checks each logical step, but the resulting work is not automatically readable to people. The Clay Mathematics Institute (CMI) has been reviewing it since its release, so the question of whether the proposed solution holds remains under examination.
Questions about where the work came from
The debate began before publication. NYU mathematician Tristan Buckmaster accused OpenAI of building on work he and Anthropic’s Levent Alpöge had done with publicly available models. OpenAI denied using Buckmaster’s data.
Camillo De Lellis of the Institute for Advanced Study says a definitive answer is probably impossible because language models are black boxes. That makes it difficult to establish exactly what information shaped a result, even as researchers argue about similarities between projects. The source does not establish that OpenAI used Buckmaster’s work; it describes an accusation and a denial.
In early October, OpenAI released roughly 400 more results as a loose collection on GitHub. Many mathematicians felt that publishing them this way cheapened their field, and the volume was difficult for the math community to keep up with. The response shows that how research is presented can become part of the dispute, alongside the substance of a potential proof.
The agent swarm’s contribution is unclear
OpenAI used about 10,000 model instances for roughly 88 hours and 130 billion tokens to develop the potential Navier-Stokes solution. Yet, according to OpenAI researcher Noam Brown, nobody actually knows what that large agent swarm contributed to the result.
That uncertainty reflects a lack of systematic comparison. Tests involving thousands of agents cost too much to run, and no one has tested how long a single agent would take to solve the same problem. It is also nearly impossible to tell what concrete advantage 10,000 agents offer over 1,000.
Brown puts less than 10 percent of the success down to the multi-agent setup. In his view, the agent system attracts a disproportionate share of the credit because it is new and flashy. He says the base model matters more: it can work through tasks over long periods and run in parallel.
Two claims still need to be separated
The proof’s mathematical status and the agents’ role are different questions. Lean can check logical steps, and the CMI is reviewing the potential proof; those facts do not show how much the agent swarm helped produce it. Likewise, uncertainty about the agents does not determine whether the proof is valid.
For now, the account leaves both questions open. Mathematicians dispute OpenAI’s description of the proof’s relationship to earlier work, while the contribution of the roughly 10,000 model instances remains unmeasured. A review of the mathematics may address the proposed solution, but the reported lack of comparison tests means the claimed value of the multi-agent setup is still difficult to assess.