OpenAI has announced what it describes as a solution to the Navier-Stokes problem, one of seven Millennium Prize Problems set out by the Clay Mathematics Institute to spotlight the toughest open questions in mathematics. The claim, however, has been quickly overshadowed by a dispute over how the breakthrough came about.
New York University mathematician Tristan Buckmaster says OpenAI sped up its own effort on the problem shortly after learning that he and a researcher at rival lab Anthropic were close to publishing their own solution.
Buckmaster’s unease was compounded by the fact that the pair had been developing their work inside OpenAI’s Codex coding model, raising the possibility that OpenAI’s team could have seen it. In a statement posted to his website, he was careful to note he wasn’t leveling an accusation, only flagging that he didn’t know what OpenAI’s system had done or whether his team’s data played any role.
OpenAI researcher Sebastien Bubeck pushed back at a press briefing, denying the company had drawn on the pair’s work or accessed material from its servers. OpenAI’s own announcement acknowledged it began the project after picking up rumours that two Millennium Prize Problems were near solved, and conceded it couldn’t fully rule out that data from the researchers’ use of its products had fed into improving its models.
Technical details released by OpenAI describe an internal system ,more capable than its latest GPT-6 Astra model — that deployed roughly 10,000 autonomous AI agents working the problem simultaneously, arriving at a solution in 88 hours. GPT-6 Astra then took about 17 hours to verify it. The Navier-Stokes equations govern how fluids like air and water move, and the long-standing open question was whether they could ever “blow up” producing physically impossible infinite speeds under certain conditions. OpenAI’s proof reportedly concludes that they can.
Bubeck called the result the high point of a year of rapid progress at the company, and OpenAI framed the effort as part of a broader mission to help scientists push research forward while keeping the public informed about how fast the technology is advancing.
The claim adds to a string of recent forays by AI labs into pure mathematics: OpenAI pointed to progress earlier this year on a decades-old problem linked to Paul Erdős, while Google DeepMind has touted its own models’ performance on competition-level maths problems. A correct solution to a Millennium Prize Problem carries a $1 million reward from the Clay Mathematics Institute, though OpenAI currently preparing for a public listing that could value it near $1 trillion — says it won’t be seeking the prize money.
Beyond the dispute with Buckmaster, the announcement lands at a moment when OpenAI is eager to shift the narrative around its safety record. The company disclosed in July that a group of its autonomous agents had breached third-party platform Hugging Face during a security test, an episode that came alongside a comparable incident at Anthropic. Both cases have fueled fresh political pressure over AI oversight, with US senators recently pushing for a permanent ban on so-called AI “superintelligence” systems capable of outperforming humans across the board.


