OpenAI Just Solved One of Math's Greatest Unsolved Problems
On a single day in September 2025, the artificial intelligence landscape shifted. OpenAI announced that one of its unreleased next-generation models produced a formal proof addressing the Navier-Stokes existence and smoothness problem — a Clay Mathematics Institute Millennium Prize problem that has remained unsolved for nearly a century. The run reportedly consumed $22 million in compute and 3 billion output tokens across just six days.
This is not a minor academic footnote. The Navier-Stokes equations govern how gases and liquids move — from the coffee swirling in your cup to the airflow over a jet wing. Whether these equations always produce smooth solutions, or whether they can "blow up" into infinite velocity within finite time, has been one of the deepest open questions in mathematics.
📅 Information as of: September 2025
Understanding the Navier-Stokes Problem
What the Equations Actually Describe
The Navier-Stokes equations model fluid motion by balancing acceleration, pressure gradients, momentum transfer, and viscosity. A key question is whether a singularity — a point where velocity diverges to infinity in finite time — can form from smooth initial conditions.
The Euler Equation Connection
When viscosity is removed entirely, the system reduces to the Euler equations, named after Leonhard Euler. Mathematicians Diego Córdoba and Luis Martínez-Zoroa had previously constructed special mathematical conditions — sometimes informally called a "devil" — that force blow-up behavior in related settings. Their work became the conceptual foundation for later progress.
For readers interested in how AI hardware and compute infrastructure are evolving to support this scale of research, see our AI laptop performance comparison guide covering the latest consumer-grade AI accelerators.
Forced vs. Unforced Stirring
A crucial distinction is between forced stirring (an external force agitating the fluid) and unforced stirring (the fluid evolving on its own). The Clay problem focuses on whether singularity formation is possible even without external forcing.
The Race Between OpenAI and Anthropic
Timeline of Events
According to public statements from mathematician Tristan Buckmaster (NYU), the sequence unfolded rapidly:
| Date | Event | Key Detail |
|---|---|---|
| Aug 15 | Euler equation partial results | Simplified case of Navier-Stokes |
| Aug 22 | Lean formal verification completed | Machine-checkable proof language |
| Sep 3 | Rumors of Anthropic progress surface | Levent Alpöge reportedly tipped OpenAI |
| Sep 6 | Three-way calls held | Anthropic employee declined participation |
| Sep 8 | OpenAI publishes results | Internal model, 10,000 agents, 88 hours |
Compute Scale and Agent Architecture
OpenAI's internal system reportedly deployed 10,000 AI agents running continuously for 88 hours. For context, that is roughly 880,000 agent-hours of parallel mathematical reasoning — a scale no human research team could replicate.
Data Provenance Questions
Buckmaster has stated he does not know whether his session data was used in training, and OpenAI has denied that researchers or agents viewed the external work. The company also noted that its proof approach differs substantially from the Euler-equation results, and that the exact theorems proved are not identical.
For a broader look at how AI-driven valuation and competition are reshaping the tech sector, see our analysis of the SpaceX IPO valuation and its $2 trillion market cap question.
What This Means for the Future of Mathematics
The Four-Minute Mile Analogy
Buckmaster compared the situation to Roger Bannister's 1954 four-minute mile: once one runner proved it was possible, others followed rapidly. OpenAI, having learned that Anthropic was making progress, reportedly deployed massive compute to close the gap — not by copying, but by following a known-viable path.
Verification Still Pending
The Clay Mathematics Institute has not yet accepted the proof. Formal verification via the Lean proof assistant is one step; peer review and institutional acceptance are separate. Whether AI-generated proofs qualify for the prize remains an open question.
The Bigger Shift
The competitive dynamic has changed. It is no longer human vs. human — it is frontier AI model vs. frontier AI model. As Buckmaster noted, the human contribution in these projects is increasingly about direction-setting and framing, not derivation.
📅 Information as of: September 2025
Key takeaways:
- AI models are now producing research-grade mathematical results
- Compute scale ($22M, 10,000 agents) is the new bottleneck
- Verification and peer review remain essential human checkpoints
- Expect more Millennium-level breakthroughs from AI systems in the coming years
