Navier-Stokes: Doubts Mount Over OpenAI’s Proof
Mathematicians challenge how OpenAI presents its potential proof of the Navier-Stokes problem. The controversy concerns both the mathematical approach and the actual role of the 10,000 agents invoked by the company.
OpenAI faces growing criticism over its claimed breakthrough on the Navier-Stokes problem after presenting a 166-page proposed solution in early September. According to several experts cited by Scientific American, the demonstration does not truly follow the approach announced by the company. The debate matters because it touches both OpenAI’s scientific credibility and how AI models are evaluated when tackling open research problems. At this stage, the mathematical community proceeds cautiously—and with some fatigue as well.
Une preuve potentielle qui divise les spécialistes
The company presented in September a potential solution to the Navier-Stokes problem, explaining it followed the path laid out by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. Experts interviewed by Scientific American dispute this reading. Gonzalo Cao-Labora, a mathematician at EPFL, believes that “When you go to the details, it is a different approach.” In other words, the similarities are mostly superficial, not structural.
The document was written in Lean, a formal verification language that checks every logical step of a reasoning. This choice allows verifying the internal coherence of a proof but does not make it readable for a human without additional mathematical translation work. The Clay Mathematics Institute is still reviewing the proposal since its publication. Nothing indicates, for now, that a definitive verdict is imminent.
Ce que la communauté reproche aussi à OpenAI
Before publication, NYU mathematician Tristan Buckmaster accused OpenAI of relying on work conducted with publicly accessible models by himself and Levent Alpöge of Anthropic. The company denied this, stating it did not use Buckmaster’s data. Camillo De Lellis, from the Institute for Advanced Study, recalls that a definitive answer is probably out of reach because language models remain black boxes. On this point, everyone sticks to their interpretation of the traces left by the systems.
The controversy is not limited to scientific attribution. In early October, OpenAI also published about 400 other results as a batch on GitHub, which irritated some mathematicians. Many saw this as a way to trivialize work they consider demanding and fragile. The gesture was perceived as another signal: AI outputs can impress, but their context matters as much as their volume.
Les 10 000 agents ont-ils vraiment changé le résultat ?
OpenAI claims to have mobilized about 10,000 model instances over 88 hours and 130 billion tokens to develop the potential solution to the Navier-Stokes problem. Yet, according to researcher Noam Brown, no one really knows what this swarm of agents contributed to the final result. Comparative experiments at this scale are too costly, and there is no test to properly measure the advantage of 10,000 agents versus 1,000 or a simpler configuration.
Noam Brown attributes less than 10% of the success to the multi-agent system. His argument is clear: multi-agent architectures attract attention because they are new, but they are not necessarily the main contributors. According to him, what matters most is the base model, capable of progressing long on a task and executing multiple calculations in parallel. In this context, the question is not only whether AI can help mathematicians but also which part of the setup truly deserves credit.
Key points
- In early September, OpenAI published a 166-page proposal.
- Gonzalo Cao-Labora speaks of a “different” approach.
- OpenAI mentions about 10,000 agents over 88 hours.
- Noam Brown attributes less than 10% to multi-agent systems.
- The Clay Mathematics Institute is still reviewing the proof.
By the numbers
- 166 pages — length of OpenAI’s proposal, early September 2026.
- 10,000 instances — volume of agents invoked by OpenAI, according to Noam Brown.
- 88 hours — computation time claimed by the company for the potential solution.
- 130 billion tokens — volume of computation data mentioned by OpenAI.
- 400 results — batch published in early October on GitHub, according to the source text.