Mathematicians Challenge OpenAI Over Navier-Stokes Proof

In early September, OpenAI released a 166-page proposed solution to the Navier-Stokes problem written in the Lean formal verification language. While the company stated that its work built upon methodologies established by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa, academic experts now challenge that assertion. Gonzalo Cao-Labora of EPFL stated that any commonalities are purely superficial, noting that "when you go to the details, it is a different approach." The Clay Mathematics Institute continues to evaluate the paper.

Controversy has also surfaced regarding the data and distribution methods behind the research. NYU mathematician Tristan Buckmaster previously alleged that OpenAI built upon prior work conducted alongside Anthropic's Levent Alpöge with open models. OpenAI denied using Buckmaster's data, though Institute for Advanced Study researcher Camillo De Lellis observed that the black-box nature of neural networks makes verification nearly impossible. Tensions heightened in early October when OpenAI uploaded roughly 400 additional outputs to GitHub, leaving mathematicians frustrated by the uncurated release.

Questions also surround the compute resources used to generate the potential proof. OpenAI deployed roughly 10,000 model instances over 88 hours, burning through approximately 130 billion tokens. Yet OpenAI researcher Noam Brown acknowledged that the team cannot determine the exact contribution of this massive multi-agent swarm. Brown estimated that the multi-agent structure was responsible for less than 10 percent of the run's success, suggesting that flashy multi-agent systems often receive excessive credit.

According to Brown, the primary driver of the result was the core model's ability to execute long-running, parallelized reasoning tasks. For practitioners, these findings indicate that simply expanding multi-agent orchestrations may yield diminishing returns compared to enhancing base model performance. Furthermore, the incident underscores ongoing friction between fast-paced AI deployment strategies and traditional mathematical peer review.