OpenAI reported that an internal, unreleased model, reportedly more powerful than GPT-6 Astra, proposed a solution to Navier-Stokes.

How the research proceeded

The company says it deployed approximately 10,000 agents that operated continuously for 88 hours to search for a formal argument and candidate steps.

• 10,000 agents ran coordinated searches across mathematical representations to propose candidate lemmas and potential proof steps over continuous computation.
• 88 hours denotes the uninterrupted compute window during which agents explored alternate constructions and refined symbolic arguments toward a consolidated demonstration.
• Company representatives describe a conditional scenario where a finite-time singularity emerges in the fluid equations, challenging classical smoothness assumptions used in analysis.

Findings and next steps

According to OpenAI, the computations indicate that under specific flow conditions a finite-time singularity may form, causing the usual mathematical description to break down.

The result tackles the Navier-Stokes existence and smoothness problem, one of the seven Millennium Problems unresolved for about 90 years.

Verification process

The proposed proof must now undergo review by the broader mathematical community, which will attempt to validate each logical step and its underpinning assumptions rigorously.

If peer review confirms it, the result would be the second solved problem among the seven Millennium Problems, after the Poincaré conjecture.

OpenAI emphasizes that the model remains internal and unreleased, and that findings have been shared with experts for independent verification and scrutiny.