Drama swirls around OpenAI’s legendary mathematical milestone

The Verge reports that OpenAI said Tuesday it solved the Navier–Stokes problem — a roughly 90-year-old Millennium Prize challenge on fluid flow — using an internal model stronger than newly released GPT-6 Astra plus 10,000 concurrent agents. Training of that internal model began August 28, OpenAI said; the company does not plan to take the $1 million prize.

One day earlier, NYU's Tristan Buckmaster published related findings with Anthropic researcher Levent Alpöge. Buckmaster says he contacted OpenAI after learning the company knew of their progress, and that OpenAI's proof used a route the pair had been pursuing with Codex and Claude. He asked whether the model had been trained on or accessed their Codex sessions; he was told the model did not look up user data, then got no answer on training.

OpenAI's public line is careful: "no specific user data was accessed," while adding that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." Sébastien Bubeck said the team did not see the pair's work until the public release and that the proofs and precise results differ. Buckmaster answered on Mastodon that OpenAI is admitting use of training data from after the pair found their result.

Signal Desk's read: this is a legitimacy crisis dressed as a math breakthrough. Solving a Millennium problem with a private model and a 10,000-agent swarm would dominate any ordinary news cycle. Instead, the story is whether OpenAI's timeline and product usage data make independent discovery look convenient. The one-day gap between Buckmaster/Alpöge's related post and OpenAI's full claim is catnip for skeptics even if the technical results diverge.

OpenAI's statement is the tell. Denying "specific user data" access while leaving de-identified derived training data on the table is lawyered language, not closure. Buckmaster's unanswered training question is the hole that matters. In a market already primed by agent-misbehavior stories, "we cannot rule out" reads as structural opacity about how customer scientific work can improve the next internal model.

Bubeck's claim that the proofs differ significantly is the right technical rebuttal if independent experts confirm it. That confirmation has to come from mathematicians, not from X threads. Until then, the political read will dominate: a frontier lab with Anthropic's rival on the other paper, a Codex-using academic team, and a prize-adjacent trophy announced at agent scale.

The likelier industry effect is procedural, not mathematical. Expect universities and funders to demand clearer terms on whether research sessions can enter training, and expect Anthropic-affiliated collaborators to treat OpenAI tooling as a disclosure risk. OpenAI declining the prize does not buy goodwill if the provenance fight stays unresolved; it only avoids a second fight with the Clay Institute.

Navier–Stokes is one of seven Millennium Prize Problems. OpenAI's internal model is described as outperforming GPT-6 Astra on mathematics benchmarks. Coverage earlier from The New York Times and Wired framed the solution claim before the credit dispute fully erupted in public posts.