OpenAI’s Navier–Stokes Millennium Prize Claim and the Ethics of LLM‑Generated Mathematics

OpenAI claims a Navier–Stokes solution using an internal LLM

OpenAI released a blog post on September 1 2026 stating that its unreleased model solved the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. The agents generated 2.7 million messages and ≈130 billion output tokens for this problem alone, with a total of 4.9 million messages and ≈300 billion tokens across all attempted problems. The final Lean‑formalized proof was completed on September 6 2026 using a second‑generation model (GPT‑6 Astra). At public API rates, the token usage would cost roughly $15 million.

Competing claim from NYU and Anthropic researchers

Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) posted a PDF on September 8 2026 describing their own year‑long effort, which relied heavily on Anthropic’s Codex (GPT‑5.6 Sol). They reported a breakthrough on August 15 2026 and accused OpenAI of potentially accessing their session data without consent. Key excerpts from Buckmaster’s statement:

“I asked whether the model had been trained on, or had access to, our sessions in Codex… I was told the model did not look up user data.”
“I am not accusing anyone of anything.”
Buckmaster’s team offered to publish their result, but OpenAI declined to list Alpöge as a co‑author, citing a competitive relationship with Anthropic.

OpenAI’s response and timeline

OpenAI’s blog explains that a rumor about two Millennium problems being solved reached them on September 1. Motivated by the rumor, they launched an automated “agent” effort that produced a resolution within 88 hours and added 17 hours of Lean verification. Their statement emphasizes:

“We did not see any of their work until it was released publicly. While unlikely, we cannot rule out that de‑identified data derived from their usage of our products helped improve our models.” OpenAI also notes that their proof differs from the NYU/Anthropic result, especially in the Euler case (forced vs. unforced).

Community reactions on Hacker News

Data‑privacy concerns dominate the discussion

@_bobm and @bob1029 ask how OpenAI could sift through billions of user sessions and whether such data could be incorporated into training pipelines.
@jwr wonders about the effectiveness of the “Improve the model for everyone” toggle and whether opting out truly prevents data use.
@vb-8448 bluntly states that any data sent to an LLM will eventually be used to improve models, regardless of terms of service.

Cost and accessibility of AI‑driven mathematics

@TrackerFF points out that solving a Millennium problem with a $15 M compute budget is currently out of reach for most researchers, potentially creating a two‑tier research ecosystem.
@keremk argues that the decision to spend such an amount is rational if a competitor might claim the prize first, framing the episode as a competitive “rumor‑driven” sprint.

Technical feasibility and model capabilities

@civvv notes that LLMs excel when large bodies of prior work exist in their training data, but doubts they can generate truly novel mathematics without human guidance.
@pietz suggests that OpenAI’s model likely did not incorporate the NYU team’s recent drafts because the work was done on GPT‑5.6, which would not have been folded into the internal model by early September.

Ethical and strategic implications

@sdcfgy recommends avoiding LLM services for confidential research, likening the situation to handing a lab notebook to an untrusted party.
@Simran‑B questions the business case: spending $15 M to win a $1 M prize seems aimed more at showcasing capability than profit.
@stn_za draws a parallel to traditional mathematics, where researchers build on each other’s published work, but highlights the speed and scale differences when AI can react to rumors instantly.

What the episode reveals about AI‑driven scientific discovery

Rumor‑driven compute allocation – The OpenAI team acted on a rumor, allocating massive resources to a single problem within days. This mirrors trends in security research where a disclosed vulnerability triggers rapid, large‑scale exploitation attempts.
Opaque data‑use policies – Even when companies claim they do not “look up” user data, the possibility that de‑identified session data influences model updates remains a gray area, raising legal and ethical questions.
Economic barrier to entry – The $10‑$15 M compute cost to attempt a Millennium problem is prohibitive for most academic groups, potentially concentrating breakthroughs in well‑funded labs.
Human‑AI collaboration vs. competition – The NYU/Anthropic team used OpenAI’s own Codex for months, yet felt sidelined when OpenAI launched a parallel effort. The incident underscores the need for clear norms around credit, data sharing, and co‑authorship when multiple parties rely on the same AI services.
Verification pipelines – OpenAI’s use of Lean for formal verification demonstrates a growing workflow: LLM generates conjectures, agents explore proof space, and theorem provers certify results. Whether this pipeline can be generalized beyond problems with extensive prior literature remains open.

Open questions for the community

How should AI providers handle user data that may contain unpublished breakthroughs?
What governance structures can ensure fair attribution when multiple teams use the same commercial LLMs?
Will the cost of AI‑driven proof search drop enough to democratize access, or will it cement a monopoly on high‑impact mathematics?
Can formal verification tools keep pace with the volume of AI‑generated proof drafts, ensuring correctness without prohibitive human effort?

This post synthesizes the original announcement, the NYU/Anthropic statement, and the most up‑voted Hacker News comments to provide a self‑contained overview of the technical, ethical, and economic dimensions of the OpenAI Navier–Stokes claim.