On September 8, 2026, NYU Courant mathematician Tristan Buckmaster published a statement: on September 6, during calls involving Sebastian Bubeck, head of OpenAI’s math initiative, he was told that an internal OpenAI model had produced a proof of “blow-up” for the Navier–Stokes equations with an external force. The proof has not been published, there is no independent verification, and the Clay Mathematics Institute still lists the Millennium Prize problem as unsolved.

image

What happened

According to Buckmaster’s statement, the proof concerns “the existence of forced blow-up in R³ and T³” for the Navier–Stokes equations with a smooth external force, i.e., variants c and d of the Clay/Fefferman formulation. The document is estimated to be about 100 pages; Buckmaster himself has not seen it. The day before the statement, on September 7, he and Levant Alpöge, an Anthropic employee, published their own LLM-assisted results on blow-up with smooth force for the Euler, Boussinesq, and incompressible porous media equations, with the first LLM-generated result from that work verified in the Lean formal verification system.

Context

The Navier–Stokes equations problem is one of the Clay Mathematics Institute’s Millennium Prize Problems, carrying a $1 million prize. The “forced blow-up” formulation relies on an external force term that some mathematicians consider a loophole in the Clay conditions, so even a correct proof may not close the Millennium Prize problem as such—that is a separate question. The announcement came amid a priority dispute: Buckmaster claims he was offered the chance to formalize the OpenAI result without Alpöge, while Bubeck called those accusations false. The mathematical background is the “forced” blow-up program started by Diego Córdoba and Luis Martínez-Zorita.

Why this matters for the industry

If OpenAI’s formulation is confirmed, it would be the first case of an AI model closing or nearly closing a $1 million Millennium Prize problem. For now, the confirmed part of the story is a specific mechanism: a human mathematician with LLMs (Claude and OpenAI Codex, including GPT-5.6 Sol and Astra) took Córdoba and Martínez-Zorita’s program to smooth force and the Euler equations in about a month, which Buckmaster himself calls a “Deep Blue–Kasparov moment” for mathematics. For the industry, this means the “LLM + formal verifier” loop yields reproducible results in differential equations research, though there is nothing to deploy or purchase: no public API, accessible model, or metrics. Builders can already construct research-agent workflows with Lean checkpoints, trackers for the verification status of high-profile scientific claims, and provenance/attribution tools; the research copilot market is at the narrative-formation stage, not the sales stage.

Why this matters for users

Declaring the “Millennium Prize problem solved” is premature: the prize has not been awarded, and the statement has become the subject of an authorship dispute, so it is too early to draw conclusions from it. What to watch specifically: publication of a preprint and Lean certificate from OpenAI, the official response from the Clay Mathematics Institute, and further statements from Bubeck and Buckmaster. The work on the Euler and Boussinesq equations is already public—see Buckmaster’s NYU page statement for details.

What is still unknown / limitations

The proof text has not been published, so its content, correctness, and connection to the Clay problem formulation are unknown. It is unclear whether the Clay Mathematics Institute will accept the external force variant (variants c and d) as a solution to the Millennium Prize problem, even if the proof turns out to be correct. The name of the model that produced the proof has not been disclosed, and there is no data on API, pricing, or metrics. The parties’ positions in the priority dispute are statements by the participants themselves and have not yet been independently verified.

Sources

Author

Look at AI, editorial team