OpenAI Faces Academic Pushback Over 166-Page Navier-Stokes Mathematical Proof
Mathematicians question OpenAI's Navier-Stokes proof, while internal researchers admit little is known about the true value of running a 10,000-agent swarm.
In early September, OpenAI released a 166-page paper offering a potential solution to the famous Navier-Stokes problem. The company stated that its work built on methods previously developed by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. However, that characterization is facing pushback from prominent figures in the mathematics community.
According to a report from Scientific American, outside researchers who analyzed the document found that the parallels between the methods are minimal. Gonzalo Cao-Labora, a mathematician at EPFL, pointed out that the similarities are only surface-level, stating, "When you go to the details, it is a different approach." The paper itself was composed in Lean, a formal verification language designed to validate each logical step of an argument computationally. While Lean ensures machine-verifiable rigor, it also produces proofs that are extremely difficult for human experts to read and interpret easily. The Clay Mathematics Institute (CMI) has been reviewing the 166-page submission since it became public.
Disputed techniques and attribution questions
Attribution questions around the release began even before the paper went live. Tristan Buckmaster, a mathematician at New York University, claimed that OpenAI had relied on work he conducted alongside Anthropic's Levent Alpöge using publicly accessible models. OpenAI denied using Buckmaster's data in the project. Camillo De Lellis, a mathematician based at the Institute for Advanced Study, observed that arriving at a definitive answer regarding data provenance is likely impossible because modern language models remain black boxes.
Tensions between OpenAI and academic researchers grew further in early October when the company released approximately 400 additional mathematical findings as a loose collection on GitHub. The sheer volume of material overwhelmed researchers trying to evaluate the proofs, leading several mathematicians to express frustration that publishing uncurated dumps cheapens the discipline.
The compute cost and the multi-agent puzzle
Beyond disputes over attribution, OpenAI's internal evaluation of the project reveals lingering questions about how the solution was produced. Generating the potential proof required a massive computational footprint: OpenAI mobilized roughly 10,000 model instances running for approximately 88 hours, consuming roughly 130 billion tokens.
Despite the scale of the deployment, OpenAI researcher Noam Brown acknowledged that the company does not know how much the multi-agent architecture actually contributed to solving the problem. Because running controlled comparison benchmarks with thousands of parallel agents is prohibitively expensive, systematic research into behavior at this scale has not been performed. As a result, the team never tested how long an individual agent would take to tackle the problem, making it nearly impossible to determine what real advantage 10,000 agents offer over 1,000 agents.
Brown estimated that the multi-agent framework accounted for less than 10 percent of the project's success. He noted that agent swarms often receive an outsized share of credit because the setups are flashy and novel. In reality, Brown argued, the critical factor is the underlying strength of the base model itself, specifically its ability to persist through difficult tasks over long durations and operate effectively in parallel.
What it means for developers
The findings shared by OpenAI's research team offer a pragmatic lesson for engineers designing automated workflows. While complex multi-agent swarms frequently capture headlines, practical performance often relies far more on the raw reasoning ability of the foundational model than on coordinating thousands of independent agents. When dealing with difficult domain-specific challenges, deploying massive swarms can dramatically increase token consumption and infrastructure costs without delivering proportional gains.
For engineering teams experimenting with automated problem-solving, comparing base model architectures before investing in complex orchestration is often the most cost-effective path forward. Developers exploring these capabilities can try top AI models cheaply through one API at https://apixoai.online to evaluate how different systems handle structured reasoning, code generation, and formal logic tasks.
Source: Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
