OpenAI Says AI Cracked a 90-Year-Old Math Problem. The Mathematicians Who Got There First Are Not Happy.

Key points

• OpenAI announced on Tuesday that an internal AI model solved the Navier-Stokes problem, a fluid-dynamics puzzle unsolved for roughly 90 years.
• The solution carries a $1 million prize, which OpenAI says it will not claim.
• NYU professor Tristan Buckmaster published findings on a closely related problem the day before, and says he worries OpenAI used his unpublished drafts, stored in OpenAI's own Codex tool, to reach the same solution.
• OpenAI says no specific user data was accessed, but concedes it cannot fully rule out that anonymised data from its products influenced the model's training.
• The internal model behind the result is more capable than GPT-6 Astra, OpenAI's newest publicly available model, and began training on 28 August.

OpenAI has announced it solved the Navier-Stokes problem, a set of equations that describe how liquids and gases move, from water through a pipe to air around an aircraft wing. Mathematicians have tried to prove these equations work perfectly under all conditions for about 90 years. It is one of seven Millennium Prize Problems, a list of the hardest unsolved questions in mathematics, each carrying a $1 million reward for a correct solution.

The company used an internal AI model, described as more capable than its newly released GPT-6 Astra, a large language model (the technology behind chatbots like ChatGPT), running alongside 10,000 AI agents working at the same time. An AI agent is software that can carry out complex, multi-step tasks on its own without constant human instruction. OpenAI says the model started training on 28 August and quickly showed strong performance on mathematics tests.

Why is this controversial?

The timing is the problem. One day before OpenAI's announcement, Tristan Buckmaster, a mathematics professor at New York University, published his own findings on a closely related version of the problem, work carried out with Levent Alpöge, a researcher at Anthropic (the company behind the Claude AI assistant).

Buckmaster says he and Alpöge stored all their working drafts inside OpenAI's Codex, a coding and reasoning tool. When he saw OpenAI's result, he was alarmed: the proof followed a route he and Alpöge had been developing. He contacted OpenAI to ask whether the company's model had been trained on, or had seen, the material he typed into Codex.

He says he received one denial but no clear answer when he pressed on the training question specifically.

OpenAI's Tuesday blog post addressed this directly, stating that "no specific user data was accessed in order to solve this problem." The company did add a notable caveat: "while unlikely, we cannot rule out that de-identified data" (that is, data with personal details stripped out) from users' product activity "helped improve our models."

Sebastien Bubeck, a senior researcher at OpenAI, said the company's proof and Buckmaster's "differ significantly" and that staff did not see the researchers' work until it became public. Buckmaster pushed back on Mastodon, arguing that OpenAI is "openly admitting they used training data from a period after we found our result."

What does this mean for ordinary people?

For most people, the immediate effect is small. Better mathematical AI could eventually speed up research in engineering, medicine, and climate modelling, but those benefits are years away.

The sharper near-term question is about trust. Millions of people type private notes, drafts, and unpublished ideas into AI tools every day. This dispute, first reported by The Verge AI, puts a spotlight on what those companies do, or might unintentionally do, with that material during model training. If you store sensitive or original work inside any AI product, it is worth reading the platform's data and training policy before you do.

OpenAI says it will not take the $1 million prize.

Common questions

Could OpenAI really have used someone's private drafts without knowing it?

Possibly, in an indirect way. Companies sometimes train models on large pools of user data that have been anonymised, meaning personal details are removed. If that process is imperfect, fragments of a user's work could influence a model without any engineer deliberately choosing to look at it.

What is a Millennium Prize Problem?

The Clay Mathematics Institute named seven famously unsolved mathematical questions in 2000 and offered $1 million for each correct solution. The Navier-Stokes problem is one of them; only one other, the Poincaré conjecture, has been solved so far.