What happened
OpenAI reports that after rumors on 1 September 2026 that other researchers had resolved Millennium problems, it launched a multi-agent evaluation of an internal model still in training and more capable than GPT-6 Astra. About 10,000 concurrent agents ran on Navier–Stokes; OpenAI says they reached a resolution on 5 September after roughly 88 hours, then Astra spent about 17 hours on Lean formalization (complete 6 September). The claimed theorem is finite-time blowup for an initially smooth fluid at rest, with a smooth force and bounded energy, matching Fefferman’s statements C and D rather than the unforced problem many analysts treat as the essential question. OpenAI says it will not claim the Clay prize. The Clay Mathematics Institute wrote on 11 September that the problem has ‘apparently been settled’ while stressing an unhurried evaluation. Nature covered the claim as a company announcement. Tristan Buckmaster alleges OpenAI moved after learning of his work with Levent Alpöge; OpenAI says the proofs differ (forced Navier–Stokes vs their forced Euler/IPM/Boussinesq program) and that it did not see their manuscripts. Hodge, Birch–Swinnerton-Dyer, Riemann, Yang–Mills, and P vs NP are not recorded as solved. This event is disputed: lab-published and Lean-accompanied, not independently reconstructed, not prize-awarded, and contested on both the forced formulation and priority.