OpenAI claims to achieve AI breakthrough in million-dollar math challenge, sparking controversy

Mathematicians question OpenAI's use of unpublished research

TLDR: OpenAI claims that approximately 10,000 AI agents cracked part of the Navier-Stokes mathematical challenge in 88 hours. The AI system generated 2.7 million messages and approximately 130 billion tokens during the mission. The capabilities of its underlying model exceed those of any OpenAI system currently open to the public. Tristan Buckmaster, a mathematician at New York University (NYU), suggested that OpenAI may have obtained details of his unpublished work. OpenAI said it will not pursue a $1 million bonus and that the certificate still needs to be independently verified.

AI breaks the Navier-Stokes problem

OpenAI announced that an experimental network of approximately 10,000 AI agents successfully solved part of the Navier-Stokes existence and smoothness challenge in 88 hours. This mathematical puzzle is one of seven "Millennium Award problems", each of which will receive a $1 million prize from the Clay Institute for Mathematics.

We are sharing solutions to the Navi Stokes Millennium Prize problem, one of the most profound problems at the forefront of mathematics.
The proof was generated by a set of agents, using the OpenAI next-generation model that is more powerful than GPT-6 Astra.
This problem...
-- OpenAI (@OpenAI) September 8, 2026

The Navier-Stokes equations are used to simulate the fluid dynamics of substances such as air and water. For a long time, the core question in mathematics has been whether smooth fluid motion will eventually deteriorate and produce infinitely increasing speeds. According to the conclusions of the OpenAI system, such a crash is possible.

The proposed proof describes a fluid vortex that stretches and strengthens until it explodes at a finite velocity for a finite time, although the total energy of the system remains limited throughout the process. Throughout the entire problem resolution process, AI agents communicated through approximately 2.7 million messages and generated approximately 130 billion tokens. Under OpenAI's standard pricing structure, the cost of such a huge computing resource is approximately $10 million.

The system is a proprietary model that OpenAI claims is superior to GPT-6 Astra, the most advanced publicly available model currently available. After the agent finished its work, Astra spent another 17 hours verifying and formalizing the certificate. According to OpenAI, training on the advanced model began in late August. By September 1, the organization received information that other parties had solved two Millennium Award challenges. This prompted OpenAI to turn its AI agents towards unresolved issues. By September 5, about 88 hours after it started, OpenAI had produced a candidate solution.

Mathematicians question use of unpublished research

The discovery immediately triggered a reaction from New York University mathematics professor Tristan Buckmaster. He revealed that he and Lvent Alpöge, a mathematician at Anthropic, had been working on very similar problems for about a year. Buckmaster disclosed that both researchers used OpenAI's Codex platform during the investigation. On September 3, he discovered that details about their research progress had been transmitted to OpenAI.

He claimed that OpenAI began handling the Navi Stokes challenge after receiving the information. He emphasized that almost no other mathematician used the same method as he and Alpöge used. OpenAI denies allegations of direct access to Buckmaster and Alpöge research material. The organization stated that no user information was retrieved and there was no conclusive evidence of contact. Still, OpenAI acknowledges that the possibility that researchers using data from OpenAI tools could help model enhancement cannot be completely ruled out.

OpenAI congratulates the two researchers, believing that their contributions occurred simultaneously rather than sequentially. The company also pointed out that the two proofs differ in methodology and actually establish slightly different mathematical results.

The proposed solution has not yet been verified by an independent review. The Clay Mathematical Institute stipulates that candidate solutions must undergo extensive peer review over several years before any bonuses can be awarded. OpenAI confirmed that it would not pursue a $1 million reward and viewed the achievement as evidence of the accelerated development of its research infrastructure.