OpenAI Says Its AI Cracked a 90-Year Math Problem. Mathematicians Aren't Convinced.

What OpenAI's system addressed is real but narrower than the viral framing suggests. The deeper, unforced versions of the problem, Clay options A and B, which ask whether singularities can arise with no external forcing at all, remain unresolved — and many mathematicians consider that the genuinely fundamental open question ^[2]. Whether OpenAI's approach can be extended to reach that harder case is, in the company's own framing, "not yet clear" ^[2].

Lean verification confirms only that the proof's logical steps are internally consistent; it cannot by itself confirm the proof meaningfully resolves the problem as mathematicians understand it, and no independent peer review has occurred ^[2][3]. The Clay Institute's own rules require a proposed solution to sit in a peer-reviewed journal for at least two years and achieve "widespread acceptance" before any prize is awarded — a bar only the Poincaré conjecture has ever cleared among all seven Millennium Prize problems since 2000 ^[1][2].

The entire sprint was set off by a claim that wasn't true. OpenAI researcher Sébastien Bubeck has said the Millennium Prize push began specifically because "viral social media rumors suggested rival lab Anthropic had resolved two Millennium Prize problems" ^[1][3]. In reality, Anthropic's Claude had improved bounds on the Riemann hypothesis using 60 subagents — a narrower result that social media conflated with the unrelated, unsolved Navier-Stokes problem ^[1].

Separately, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had spent roughly a year working on related fluid-dynamics questions and, on August 15, 2026, proved finite-time blowup for the 3D Euler equations ^[1][2][3]. Buckmaster alleges that during September 6 negotiations, Bubeck proposed a joint release that would have dropped Alpöge as a co-author because of his employment at Anthropic — an account Bubeck has called "false and inflammatory" ^[1]. Both researchers have also raised the question of whether OpenAI's agents had access to their unpublished work or Codex sessions; OpenAI says its researchers and agents did not see that work before its public release and denies inspecting private user data, while acknowledging that de-identified platform usage broadly could have helped train its models ^[1][3].

UNVERIFIED. The claim that OpenAI's AI "solved" a 90-year-old Millennium Prize problem overstates a real but partial result: the system addressed an easier, forced variant of Navier-Stokes, not the core unforced problem mathematicians still consider open, and the manuscript has not been independently peer-reviewed ^[2][3]. The rumor about Anthropic that triggered the entire race was itself false ^[1]. And the dispute over whether OpenAI raced past Buckmaster and Alpöge's unpublished work, or tried to push Alpöge off the credit line because of his Anthropic ties, remains a direct factual disagreement between the parties with no independent account yet settling it ^[1].