OpenAI states that a secret, testing-phase artificial intelligence setup cracked one of math’s most famous unsolved questions after roughly 10,000 agents collaborated for a mere 88 hours.
The organization released a proposed solution for the Navier-Stokes existence and smoothness problem, which stands as one of seven Millennium Prize Problems carrying a $1 million bounty. These formulas explain the motion of fluids like air and water, while the open question inquires whether a completely smooth flow can eventually break down and achieve speeds that increase infinitely.
The system from OpenAI argues that it can. Its mathematical proof outlined a fluid vortex that stretches and condenses further until its velocity blows up within a finite timeframe, even while the overall energy of the system remains finite.
Actual fluids cannot travel at infinite speeds under our current comprehension of physics. At sufficiently extreme levels, the fundamental premise behind Navier-Stokes that treats a fluid as a continuous medium ceases to align with physical reality.
This outcome stemmed from an internal model that OpenAI claims is vastly more powerful than GPT-6 Astra, its top publicly available model. Approximately 10,000 agents exchanged 2.7 million messages and produced roughly 130 billion tokens while tackling Navier-Stokes, before Astra spent an additional 17 hours formalizing and verifying the proof.
The practical consequences of the proof itself, however, will take time to manifest. Navier-Stokes equations see application in fields like aircraft engineering, meteorology, and blood-flow research, meaning a better grasp of where these formulas might break down could eventually enhance how researchers simulate extreme and turbulent flows.
The larger significance may lie in the framework that uncovered it. If thousands of AI agents can take on a 90-year-old mathematical hurdle in a matter of days, this identical strategy could eventually be pointed toward difficult problems within materials science, energy, aerospace, medicine, and additional domains.
OpenAI challenged
Controversy already surrounds the question of how independently this breakthrough was achieved.
New York University mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge were independently investigating a heavily connected fluid-dynamics challenge and had utilized AI instruments during their studies. Buckmaster has raised doubts over whether unreleased data originating from two mathematicians’ related investigations might have influenced OpenAI’s model, as outlined in a sequence of posts on the social platform Mastodon.
Buckmaster mentioned that he and Alpöge had spent about a year pursuing an unconventional pathway toward the issue involving a smooth external force, with Codex serving as one of the AI tools heavily utilized throughout their project.
He added that almost no one else he knew of was exploring that exact strategy and questioned how OpenAI managed to reach it just days after discovering the pair had achieved momentum.
Buckmaster recounted that during discussions with OpenAI, he was initially presented with a prompt and informed that the corporation’s internal research initiative kicked off entirely without human intervention aside from the initial problem description.
That narrative shifted as the dialogue progressed, according to Buckmaster, with OpenAI eventually admitting that the initial prompt had only been dispatched over the preceding days, subsequent to information about the work of Buckmaster and Alpöge reaching the enterprise.
OpenAI countered by stating that neither its researchers nor its agents reviewed their work prior to publication and that no particular user information was accessed. The organization noted it could not completely rule out that de-identified data stemming from their application of OpenAI products might have assisted in refining its models, while maintaining that the proofs themselves show substantial differences.
Consequently, the proof is not yet recognized as an official solution. The sponsoring Clay Mathematics Institute mandates that proposed answers endure prolonged evaluation and achieve wide consensus among mathematicians prior to considering the distribution of the $1 million award.
In the interim, OpenAI states it has no plans to claim the financial prize, characterizing the outcome instead as proof of how rapidly its research infrastructure continues to advance.
Originally published at https://www.coindesk.com/tech/2026/09/09/openai-says-10-000-ai-agents-solved-a-usd1-million-math-problem-now-mathematicians-are-fighting.
