OpenAI Claims 88-Hour Solution to Navier-Stokes Problem Amid Plagiarism Allegations

OpenAI claims it solved the 90-year-old Navier-Stokes mathematics problem in just 88 hours using an unreleased internal AI model. [2][5] The company initiated training for a new internal model in late August after determining it was significantly more capable at mathematics than its prior systems. [2] On September 1, after hearing rumors that two other Millennium Prize Problems had been solved, OpenAI decided to test its model on the remaining challenges. [2]

The company deployed around 10,000 AI agents divided into groups, with some tasked to find a proof and others instructed to look for a counterexample. [2] During this 88-hour effort, the agents generated approximately 130 billion output tokens, exchanged around 2.7 million messages, and shared useful findings with each other. [2] The initial result was subsequently formalized and verified over an additional 17 hours using the GPT-6 Astra model. [2]

The Navier-Stokes equations describe the motion of fluids such as water and air, and are utilized by scientists for modeling aerodynamics and blood flow. [1][2] The problem is one of six remaining unsolved Millennium Prize Problems, which carry a $1 million reward for a complete solution from the Clay Mathematics Institute. [1][3] The problem specifically asks whether smooth, three-dimensional fluid flows always maintain smooth behavior or if they can develop a singularity with infinite velocity. [3]

OpenAI's reported result addresses a narrower path by demonstrating a finite-time blowup for forced Navier-Stokes equations in R³ and T³. [6] This result shows that a smooth external force drives the equations toward a singularity within a finite period. [6] While OpenAI says it will not claim the $1 million prize, mathematicians still need to inspect the proof and establish if the result satisfies the institute’s requirements. [2][6]

New York University mathematics professor Tristan Buckmaster claimed OpenAI's approach matched niche, private work he developed with Levent Alpöge, a researcher at OpenAI competitor Anthropic. [1][6] Buckmaster and Alpöge had spent roughly a year studying the related Euler and Boussinesq equations using AI assistants like OpenAI Codex and Anthropic's Claude. [6] Buckmaster stated that he and Alpöge were finalizing their preliminary findings when they learned information about their progress had been passed to OpenAI. [4]

OpenAI scientists tried to convince Buckmaster to jointly announce the solution, after he contacted the company to ask when it began its research and what data its model was trained on. [3][5] Buckmaster alleged that OpenAI sent its first prompt to work on the problem only after information about his work reached the company in the past few days. [3][4] According to Buckmaster, it emerged that an entire team had been working on the problem and that an insane amount of compute had been used. [4]

Buckmaster claimed an OpenAI researcher warned he might ruin his career if he refused to collaborate and instead went public with his version of events. [3][5] Sebastian Bubeck, who leads mathematical research at OpenAI, posted that those claims are false and inflammatory. [4] Bubeck stated that he came into the discussion following academic norms and that academic standards remain of the highest importance to him. [4]

Despite this denial, Queen Mary University of London mathematics professor Abhishek Saha stated that OpenAI engaged in actions that mathematicians generally will not do. [5] The ensuing controversy has surfaced allegations of scooping, spying, and flagrant violations of long-standing academic norms that researchers fear could chill the field. [5]

The deployment of 10,000 agents and 130 billion tokens demonstrates a fundamental shift in mathematical research from individual human insight to massive, compute-driven brute-force exploration. By rapidly overwhelming a specific domain after hearing rumors of rival progress, OpenAI showed how massive scale can be weaponized to upend traditional academic timelines. Buckmaster’s own use of Claude and Codex alongside OpenAI's swarm of agents illustrates that AI tools are now deeply embedded on both sides of frontier mathematical disputes. Furthermore, the incident highlights a severe culture clash between the tech industry's race for compute dominance and academia's careful norms around research priority and attribution. What the sources don't address: Whether the formal proof generated by GPT-6 Astra actually withstands rigorous peer review by human mathematicians to be accepted as a valid finite-time blowup result.