On September 8, OpenAI announced that an unreleased AI model—more powerful than the just-launched GPT-6 Astra—solved one of mathematics' hardest open problems.
The company deployed roughly 10,000 coordinating agents that ran for 88 hours, exchanged 2.7 million messages, and consumed 130 billion output tokens. Conservative estimates put the compute cost at $10 million. The problem was the Navier-Stokes existence question, part of a set of seven Millennium Prize Problems each carrying a $1 million award. Only one had been solved before.
The model you can't rent
OpenAI described the internal model as significantly more capable than Astra, which itself ranks as the strongest reasoning model the public can access today. Training began August 28 and continues. There is no release date and no API.
That gap matters. A single math proof required millions in compute and a private model unavailable to researchers, startups, or anyone outside OpenAI's walls.
A controversy over credit
The result also sparked a fight. NYU mathematician Tristan Buckmaster says he and Anthropic researcher Levent Alpöge worked for a year on related problems using Claude and Codex—OpenAI's own tools. When rumors of their breakthrough circulated, OpenAI launched its agent swarm on September 1. Five days later, it had a proof.
Buckmaster says OpenAI never answered whether Codex session data helped train the model. OpenAI says it "cannot rule out" that anonymous usage data from its products improved later systems, but insists no specific session was accessed.
What just became normal
The technical milestone is real. The financial one is bigger: solving hard problems now requires compute budgets most institutions on Earth cannot match. When only a few labs can afford to run the models that matter, renting becomes the only option—and the landlord sets the terms.