OpenAI says one of its unreleased AI models has produced a proposed solution to the Navier-Stokes existence and smoothness problem, one of mathematics’ seven famous Millennium Prize Problems.
The model behind the result is not GPT-6 Astra. OpenAI says it used an internal system that is “significantly more capable” than Astra, coordinating thousands of AI agents to work on the problem at the same time.
The announcement immediately drew attention for two reasons. First, the mathematical claim itself is enormous. Second, the timing created a dispute over research credit after NYU mathematician Tristan Buckmaster questioned whether OpenAI may have indirectly benefited from drafts that he and Anthropic researcher Levent Alpöge had previously worked on with Codex.
OpenAI Says Its AI Found a Navier-Stokes Singularity
The Navier-Stokes equations describe the motion of fluids and appear across fields including engineering, climate science, aerospace, ocean modeling and medical research. The Millennium Prize question asks whether smooth three-dimensional solutions always remain smooth or whether the equations can develop a singularity in finite time.
OpenAI says its internal model found a construction showing that a smooth fluid can develop such a singularity under a smooth external force. According to the company, the result establishes part of the official formulation of the Navier-Stokes problem and provides a concrete mathematical path toward resolving the long-running question.
The claim is significant, but it should not yet be described as an officially awarded Millennium Prize solution. The Clay Mathematics Institute requires any proposed solution to undergo publication, scrutiny and broad acceptance within the mathematics community before it can be formally recognized.
Around 10,000 AI Agents Worked on the Problem
OpenAI says the effort involved roughly 10,000 AI agents operating in parallel. Rather than relying on a single model conversation, the company split the work across large numbers of agents that explored different approaches, tested intermediate ideas and shared useful results.
The agents reportedly spent around 88 hours reaching the proposed solution. OpenAI then used GPT-6 Astra for an additional formalization and verification stage, including work in Lean, a theorem-proving system used to check mathematical arguments more rigorously.
The scale of the experiment is striking. OpenAI said the broader effort involved millions of agent messages and hundreds of billions of generated tokens across several mathematical problems. That makes this less like a normal AI prompt and more like a large-scale computational research operation.
The Compute Cost Reached Millions of Dollars
The project was not cheap. Reports surrounding the experiment estimate that OpenAI spent millions of dollars in compute while running the agents and processing the resulting mathematical work.
That figure matters because it shows how frontier AI research is beginning to combine model intelligence with sheer computational scale. A difficult problem can now be attacked by thousands of parallel reasoning systems rather than one researcher, one workstation or one model session at a time.
Sam Altman described the result as one of the most remarkable moments in OpenAI’s history, reflecting how seriously the company views the breakthrough.
The Internal Model Is More Capable Than GPT-6 Astra
One of the most notable details in OpenAI’s announcement is that the discovery was not made by its newest public model.
OpenAI says the internal model behind the Navier-Stokes result is significantly more capable than GPT-6 Astra and is still undergoing development. That suggests the company already has systems internally that outperform what users can currently access.
For researchers watching the pace of AI progress, that may be almost as important as the proof itself. Public models are improving quickly, but internal research systems may already be operating at a noticeably higher level in mathematics, science and long-horizon reasoning.
Tristan Buckmaster Questions the Timing
The story became more complicated after Tristan Buckmaster raised questions about the timeline.
Buckmaster and Anthropic researcher Levent Alpöge had spent roughly a year working on a related mathematical route involving the Euler equations. Their research also made extensive use of AI tools, including Codex, and they released partial results shortly before OpenAI published its own work.
Buckmaster later argued that OpenAI only began aggressively pursuing the problem after learning that his team was making progress. He also questioned whether drafts that had passed through Codex could have influenced OpenAI’s model development in some indirect way.
The issue does not necessarily mean OpenAI copied the researchers’ work. The concern is more subtle: whether privately submitted research material might have contributed to model improvement before another OpenAI system later attacked a closely related problem.
OpenAI Says It Did Not See Their Research
OpenAI rejects the suggestion that its research team or AI agents directly accessed Buckmaster and Alpöge’s unpublished work.
The company says no specific user data was accessed during the Navier-Stokes project and that its researchers did not see the pair’s work before it became public.
OpenAI has also acknowledged an important limitation. It cannot completely rule out the possibility that de-identified product usage data may have contributed more broadly to improving its models.
That distinction is likely to become increasingly important as scientists, developers and academics use commercial AI systems for unpublished research. A company may not directly retrieve a private document, but researchers may still want clearer guarantees about how their interactions contribute to future model training or improvement.
The Credit Dispute Raises Bigger Questions About AI Research
The argument over priority has started to overshadow the mathematics, but it points to a problem that AI-assisted research will face more often.
Researchers increasingly use AI systems to test theories, draft proofs, write code and explore ideas before publication. Those same AI providers are also building their own models to conduct independent research.
That creates a difficult trust issue. Scientists may hesitate to share highly valuable unpublished ideas with hosted AI platforms if they cannot clearly understand how those interactions are stored, processed or used later.
The Navier-Stokes dispute could therefore become an early example of a much broader debate over intellectual contribution, training data and research confidentiality in the age of frontier AI.
Formal Verification Strengthens OpenAI’s Claim
OpenAI did not stop at publishing a conventional mathematical argument. The company also says it formalized and checked the result using Lean.
Formal verification matters because theorem-proving software can detect logical gaps that may be difficult to catch through normal peer review alone. It does not automatically guarantee that the entire scientific interpretation is correct, but it gives mathematicians another layer of evidence when evaluating the work.
This combination of AI-generated discovery and machine-checked verification could become increasingly common. Future mathematical research may involve AI systems proposing ideas while formal tools independently check whether the reasoning holds together.
The Millennium Prize Is Not Yet Officially Won
Despite the scale of OpenAI’s announcement, the Navier-Stokes Millennium Prize should not yet be described as formally awarded.
The Clay Mathematics Institute has a defined process for reviewing proposed solutions. The work must appear in an appropriate publication, remain under scrutiny for a significant period and gain general acceptance among experts before the institute considers awarding the $1 million prize.
OpenAI has also indicated that it does not intend to claim the prize itself.
The next stage will therefore happen outside OpenAI. Mathematicians will need to review the proof carefully, test its assumptions, compare it with existing work and determine whether the claimed result holds up.
Why the OpenAI Navier-Stokes Breakthrough Matters
Even with the unresolved credit dispute and the need for independent verification, the capability demonstrated by the project is difficult to ignore.
OpenAI says thousands of autonomous agents collaborated on one of mathematics’ most difficult open questions and generated a proposed solution within days. Another AI system then helped formalize that work.
That is a very different picture of AI from the familiar chatbot model.
Instead of simply answering questions, frontier systems are beginning to operate as large-scale research networks capable of exploring enormous problem spaces in parallel.
The biggest implication may not be whether OpenAI ultimately receives recognition for solving Navier-Stokes. The bigger story is what this kind of AI research infrastructure could do next.
If an unreleased model already surpasses GPT-6 Astra by a meaningful margin, expectations for what AI can achieve in mathematics and scientific discovery during the rest of 2026 may need another reset.
Sources
OpenAI — On the Navier-Stokes Millennium Prize Problem
https://openai.com/index/navier-stokes-solution/
Clay Mathematics Institute — Millennium Prize Problems
https://www.claymath.org/millennium-problems/
