OpenAI's Navier–Stokes Solution: 10,000 Agents, 88 Hours, Lean Formalization, and the Credit Controversy

Executive Summary

On September 8, 2026, OpenAI announced what may be the most significant demonstration of AI capability in scientific reasoning to date: an internal model — described as "significantly more capable than GPT-6 Astra" — coordinated a swarm of approximately 10,000 concurrent AI agents to produce a solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems that has remained unsolved for roughly 90 years.

The proof shows that the Navier–Stokes equations for three-dimensional incompressible fluid flow can develop a singularity in finite time — meaning a fluid starting from smooth, rest conditions can evolve to infinite velocity under the equations' own dynamics, despite the presence of viscosity. This resolves statements "C" and "D" in the official Clay Mathematics Institute formulation. The analytical proof spans approximately 165 pages, and OpenAI has released both the writeup and a Lean 4 formalization for independent verification.

The announcement immediately triggered a credit controversy. NYU mathematician Tristan Buckmaster alleged that OpenAI accelerated its work after learning of his and Levent Alpöge's (an Anthropic employee) parallel research on related Euler equations using AI models. OpenAI denied using their work, though acknowledged the methodological similarity. The episode has sparked a broader debate about the changing nature of mathematical research in the age of AI, with Fields Medalist Terence Tao lamenting the "decoupling between getting answers and getting understanding."

---

1. The Navier–Stokes Problem: Why It Matters

1.1 The Equations

The Navier–Stokes equations use Newton's second law of motion (F=ma) to describe how fluids move, treating the fluid as a continuous medium rather than tracking individual molecules. These equations are fundamental to:

• Aircraft design — understanding airflow over wings and fuselage
• Weather forecasting — modeling atmospheric circulation
• Blood flow analysis — cardiovascular research and medical device design
• Oceanography — current prediction and climate modeling
• Engineering — pipe flow, turbine design, combustion

The equations date to the 19th-century work of Claude-Louis Navier and George Gabriel Stokes. In 1934, Jean Leray proved that solutions exist in a generalized sense, but whether they always remain smooth became a central unanswered question.

1.2 The Millennium Prize Formulation

In 2000, the Clay Mathematics Institute named the Navier–Stokes existence and smoothness problem one of seven Millennium Prize Problems, each carrying a $1 million prize. The specific question:

Can the Navier–Stokes equations for a three-dimensional incompressible fluid with constant density develop a "singularity," even when the motion starts smoothly?

A singularity means the dynamics lead to speeds in the fluid growing without bound within a finite amount of time. This would happen despite viscosity, which tends to smooth out motion. Since a real fluid cannot move infinitely fast, this would mark a breakdown in how the equations model the fluid — requiring a shift to tracking individual particles.

The Clay formulation presents four variants:

Variant	Statement	Implication
A	Smooth solutions always exist (no blowup)	Equations are well-posed globally
B	Smooth solutions always exist for large initial data	Stronger version of A
C	Blowup can occur with smooth forcing	Equations can break down
D	Blowup can occur without external forcing	Most dramatic breakdown

OpenAI's proof establishes statements C and D — showing that blowup is possible.

1.3 Why This Is Hard

The technical challenge is that the equations must develop the breakdown through the motion of the fluid itself, not through an infinite force applied by hand. The terms describing acceleration, pressure gradients, momentum transfer, and viscosity must all become large yet cancel in a precise way, leaving a smooth external force even as velocity grows without bound.

As Terence Tao (UCLA) put it: "Is there some initial fluid configuration where water splashes around and, at some point, all the energy implodes and sends all the water particles off into infinite velocity?"

---

2. The Solution: What OpenAI Found

2.1 The Result

OpenAI's system produced an analytical proof and Lean formalization showing that an initially smooth fluid at rest can develop a singularity in finite time. The fluid has a smooth force applied to it, and its energy remains finite through the entire dynamics — from rest to singularity formation.

The solution takes the form of a vortex — a spinning swirl of fluid that spirals inward and becomes increasingly elongated, "like spaghetti." The central region shrinks while it speeds up in such a way that its energy stays finite, as required by the laws of physics.

2.2 The Proof Structure

The 165-page analytical proof demonstrates that:

1. A specific class of vortex solutions can be constructed
2. These solutions spiral inward with increasing angular velocity
3. The axial stretching compensates for viscous dissipation
4. The nonlinear terms (acceleration, pressure gradients, momentum transfer) balance precisely
5. The external force remains smooth throughout
6. Velocity grows without bound in finite time while energy remains finite

The proof was accompanied by a Lean 4 formalization — a machine-checked proof that can be independently verified by anyone with the Lean theorem prover. This is significant because AI-generated proofs have historically been difficult to verify due to their complexity and the potential for subtle errors.

2.3 The GitHub Repository

OpenAI released the formal proofs at github.com/openai/NavierStokesAndEuler, containing:

• Lean 4 formalizations of the Navier–Stokes finite-time blowup result
• Lean 4 formalizations of the Euler equation finite-time blowup result
• Build instructions for independent verification

# Clone and verify the formal proofs
git clone https://github.com/openai/NavierStokesAndEuler
cd NavierStokesAndEuler
# Follow build instructions to run Lean 4 verification

---

3. The Multi-Agent System: How It Was Done

3.1 The Architecture

OpenAI used a system of coordinating agents powered by their internal model. The architecture had several key features:

Component	Description
Internal model	Significantly more capable than GPT-6 Astra; training ongoing since August 28
Agent count	~10,000 concurrent agents for Navier–Stokes; ~100 agents for Euler
Tools	Cached internet access, code execution, mathematical computation
Communication	Agents subdivided into groups with intra-group communication
Coordination	Codex used to consolidate insights across agent groups
Safeguards	Standard frontier model evaluation monitoring and isolation

3.2 The Workflow

The effort followed this timeline:

Date	Event
August 28	Training begins on new internal model with unprecedented math performance
September 1	Rumors heard of two Millennium Prize problems being resolved; effort launched
September 1–5	Agents explore all Millennium Prize problems and easier variants
~September 3	Euler regularity disproof found by ~100 agents in ~50 hours
September 3–5	Resources shifted to Navier–Stokes; agents prompted with Euler resolution
September 5	Navier–Stokes solution found, ~88 hours after first agents launched
September 5–6	Lean formalization and verification via GPT-6 Astra (~17 hours)
September 6	OpenAI contacts Buckmaster/Alpöge; discovers their Euler work
September 8	Public announcement

3.3 Scale of Compute

The numbers are staggering:

Metric	Value
Total messages (all problems)	4.9 million
Total output tokens (all problems)	~300 billion
Navier–Stokes messages	2.7 million
Navier–Stokes output tokens	~130 billion
Euler agents	~100
Euler duration	~50 hours
Navier–Stokes agents	~10,000
Navier–Stokes duration	~88 hours
Lean verification	17 hours (GPT-6 Astra)
Estimated cost	~$15 million (per New Scientist)

This represents the largest single-task deployment of AI agents for scientific research to date.

---

4. The Credit Controversy

4.1 Buckmaster's Allegation

On September 7, the night before OpenAI's announcement, Tristan Buckmaster (NYU Courant Institute) posted on social media that he and Levent Alpöge (a mathematician employed by Anthropic) had solved the forced Euler equations — widely seen as a stepping stone to Navier–Stokes.

In a formal statement released September 8, Buckmaster alleged:

"The route to the Clay problem through a smooth force, options c and d in Fefferman's statement of the problem, is the route Luis [Martínez-Zoroa] and Diego [Córdoba] opened and the one Levent and I had quietly chosen to attack. Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement."

Buckmaster's specific concerns:

1. Timing: OpenAI heard rumors of their work on September 1 and launched the effort immediately
2. Methodological similarity: OpenAI's Navier–Stokes proof followed the same "forcing" method that Buckmaster and Alpöge had been using
3. Resource disparity: OpenAI could deploy 10,000 agents on-demand; independent mathematicians cannot
4. Potential data contamination: Buckmaster's Codex prompts over two months preceding the announcement could theoretically have influenced OpenAI's training

4.2 OpenAI's Response

At a press briefing on September 8, OpenAI mathematician Sébastien Bubeck addressed the allegations:

Claim	OpenAI's Response
Used Buckmaster/Alpöge's prompts or proofs	"We did not use their prompt or proofs to prompt our models"
Euler solution was influenced	Euler was solved by "totally different means"
Navier–Stokes method was copied	Methodological similarity acknowledged but denied influence
Training data contamination	Investigation confirmed Buckmaster's Codex prompts "could not have influenced the system in any way"
Timeline	Proof developed over the weekend, after rumors were heard

OpenAI's blog post included a nuanced acknowledgment:

"While unlikely, we cannot rule out that de-identified data derived from [Buckmaster and Alpöge's] usage of our products helped improve our models."

4.3 The Euler Distinction

An important technical distinction emerged:

Result	OpenAI	Buckmaster & Alpöge
Euler equations	Unforced version (no external force)	Forced version (with external force)
Navier–Stokes	Full proof (statements C and D)	Partial — related method, not complete proof
Method	Similar "forcing" approach for Navier–Stokes	Different approach for Euler
AI models used	Internal OpenAI model (unnamed)	Anthropic Claude + OpenAI Codex + Astra

OpenAI offered Buckmaster and Alpöge visibility into all prompts used and the proof itself, and recognized the priority of their Euler work.

4.4 The Broader Debate

The controversy has sparked a fundamental debate about mathematical research in the AI era:

Terence Tao (UCLA, Fields Medalist) lamented:

"There's been this very strange and unprecedented decoupling, this year alone, between getting answers and getting understanding."

Tao's broader concern: the ability of AI labs to deploy massive, on-demand compute the instant they hear a rumor is changing incentives for how openly researchers share promising work. A mathematician who publishes a promising approach may find their idea immediately scaled by a well-funded lab with thousands of agents.

Luis Martínez-Zoroa (CUNEF University, Madrid), who co-developed the "forcing" approach with Diego Córdoba, said: "We're a little bit in shock. If it's done, that will be a big surprise for us."

Martin Bridson (Clay Mathematics Institute) called it "an exciting day, as we contemplate the announcement of major advances in the human understanding of mathematics."

---

5. The Lean Formalization: Machine-Checked Mathematics

5.1 Why Lean Matters

The Lean theorem prover is a formal verification system that requires every step of a mathematical proof to be explicitly justified according to the rules of logic. Unlike a traditional mathematical paper, where readers must trust the author's reasoning, a Lean proof can be independently verified by a computer.

This is particularly important for AI-generated proofs because:

1. Complexity: The 165-page proof may contain steps that are difficult for humans to verify manually
2. Trust: AI systems can produce plausible-sounding but incorrect reasoning
3. Reproducibility: Anyone with Lean 4 can verify the proof without trusting OpenAI

5.2 The Verification Process

The Lean formalization was produced by GPT-6 Astra over 17 hours following the agent swarm's analytical proof. The process involved:

1. Translating the analytical proof into Lean 4 syntax
2. Filling in any gaps where the analytical proof was informal
3. Running the Lean compiler to verify every logical step
4. Producing machine-checkable proof certificates

5.3 Community Access

The formal proofs are available at github.com/openai/NavierStokesAndEuler. The repository includes:

• Lean 4 source files for both Navier–Stokes and Euler results
• Build instructions for local verification
• Dependencies and environment setup guidance

This represents a new model for mathematical publication: AI-generated proof + machine-checked formalization + open repository.

---

6. Connection to the Frontier Model Landscape

6.1 The Internal Model

OpenAI described the model used as "significantly more capable than GPT-6 Astra." This is notable because:

• GPT-6 Astra was announced September 3, 2026, as the first model to trigger OpenAI's "critical" cybersecurity threshold (Ai News Week 2026 09 01 2026 09 07)
• The internal model used for Navier–Stokes was trained starting August 28 and was still undergoing training at the time of the announcement
• Its performance "continues to improve"

This suggests OpenAI is operating with models that are more capable than their public releases — a pattern consistent with the frontier model development cycle.

6.2 Comparison with Other AI Math Efforts

Organization	Approach	Result	Scale
OpenAI	10,000 agents, internal model	Navier–Stokes proof + Lean	~130B tokens, 88 hours
Buckmaster & Alpöge	Human + Claude + Codex + Astra	Forced Euler solution	Human-led, AI-assisted
Anima Anandkumar (Caltech)	Physics-informed neural network	Euler zero-viscosity solution	Neural network approach
Anthropic	Claude models (used by Buckmaster/Alpöge)	Euler equations (via researchers)	Researcher-driven

6.3 The Multi-Agent Paradigm

This effort demonstrates a new paradigm for AI research: massive multi-agent coordination for complex problem-solving. Key observations:

• Specialization: Agents were subdivided into groups with different problem variants
• Cross-pollination: Codex consolidated insights across groups
• Iteration: Model updates were applied mid-effort as training continued
• Scale: 10,000 concurrent agents represents an order of magnitude beyond previous multi-agent experiments

This connects to the broader trend of agentic systems moving from single-agent tool use to large-scale coordination, as seen in Anthropic's MHS for physical equipment (Anthropic Model Hardware Standard Mhs Research Preview Ai Agents Physical Lab Equipment 2026 09 09) and Arm's Total Design for Physical AI (Arm Physical Ai Ecosystem Launch Total Design Robotics Capability Framework Css Mobile 2 Ai Portal 2026 09 10).

---

7. Implications for Scientific Research

7.1 The Resource Disparity

The most unsettling implication is the resource gap between AI labs and independent researchers:

• OpenAI deployed ~10,000 agents for 88 hours, consuming ~130 billion tokens
• Estimated cost: ~$15 million
• Independent mathematicians work with human-scale resources and open-source tools
• The ability to "throw compute at a problem" changes the nature of mathematical competition

As the AAAS reported: "For independent mathematicians, the disparity in resources is becoming harder to ignore."

7.2 The Open Science Question

The episode raises questions about open science in the AI era:

1. Should promising research directions be published immediately, knowing AI labs may scale them?
2. How do we attribute credit when AI systems may have been trained on publicly available work?
3. What role should human mathematicians play when AI can produce proofs independently?
4. How do we maintain the culture of collaboration when resource disparities create incentives to withhold information?

7.3 The Understanding Gap

Terence Tao's observation about the "decoupling between getting answers and getting understanding" is perhaps the most profound implication. AI systems can produce correct proofs without necessarily providing the insight that makes mathematics valuable — the intuition, the connections to other areas, the new techniques that enable future work.

The 165-page proof, while machine-verified, may not immediately yield new mathematical tools or perspectives for the community.

---

8. What Happens Next

8.1 Verification Timeline

The mathematical community will need time to:

1. Review the analytical proof — 165 pages of advanced PDE analysis
2. Run the Lean formalization — independent machine verification
3. Compare with Buckmaster/Alpöge's work — determine the relationship between the approaches
4. Assess the Euler results — both OpenAI's unforced and Buckmaster/Alpöge's forced versions

This process could take weeks or months. The Clay Mathematics Institute will need to evaluate whether the proof meets the criteria for the Millennium Prize, though OpenAI has stated it does not intend to claim the prize.

8.2 The Internal Model

OpenAI indicated the internal model's training is ongoing and its performance "continues to improve." The company stated:

"We are focusing on understanding this model, and using what we learn to help us guide and pace how we pursue further advances in capability."

This suggests a period of deliberate pacing rather than immediate public release — consistent with OpenAI's stated goals of building "steerable, accountable, and connected" AI systems.

8.3 The Remaining Millennium Problems

OpenAI's system attempted all six unsolved Millennium Prize problems. The results for the others were not announced, suggesting they were either not resolved or not yet verified. The remaining problems are:

Problem	Status
Poincaré conjecture	Solved (Perelman, 2003) — not among the six unsolved
Riemann hypothesis	Still open; attempted by OpenAI agents
P vs NP	Still open; attempted by OpenAI agents
Yang–Mills existence and mass gap	Still open; attempted by OpenAI agents
Hodge conjecture	Still open; attempted by OpenAI agents
Birch and Swinnerton-Dyer conjecture	Still open; attempted by OpenAI agents

---

9. Key Takeaways

1. AI has solved a Millennium Prize Problem. Whether one accepts OpenAI's claims about independence, the result itself — a proof of finite-time blowup for Navier–Stokes, formalized in Lean — represents a genuine advance in mathematical knowledge.

2. The multi-agent scale is unprecedented. ~10,000 concurrent agents, 2.7 million messages, 130 billion tokens for a single problem represents a new order of magnitude in AI-assisted research.

3. Lean formalization is essential for trust. The machine-checked proof provides a level of verification that traditional mathematical publication cannot match, particularly for AI-generated results.

4. The credit controversy is systemic, not incidental. The resource disparity between AI labs and independent researchers, combined with the ability to rapidly scale promising approaches, creates structural incentives that may undermine open scientific collaboration.

5. The understanding gap is real. Producing a correct proof is not the same as producing mathematical insight. The community needs to grapple with what it means when answers come without understanding.

6. This is a snapshot, not a culmination. OpenAI explicitly framed this as evidence of "the next period of AI progress," not an endpoint. The internal model is still training, and performance is still improving.

7. The mathematical community is at an inflection point. As Tao observed, the dynamics of how mathematical research is conducted are fundamentally changing. The community needs new norms, new attribution models, and new ways to ensure that AI serves mathematics rather than replacing it.

---

10. References & Resources

Official Sources

• On the Navier–Stokes Millennium Prize Problem — OpenAI announcement, September 8, 2026
• openai/NavierStokesAndEuler — Lean 4 formal proofs on GitHub
• Navier–Stokes Millennium Prize Problem — Clay Mathematics Institute problem statement
• Buckmaster Statement — Tristan Buckmaster's formal statement, September 8, 2026
• Alpöge & Buckmaster: Euler Equations — Forced Euler equations paper, September 7, 2026
• Anandkumar et al.: Euler Equations — Caltech physics-informed neural network approach

Coverage (Selected)

• OpenAI claims to have solved the 90-year-old Navier-Stokes math problem — CNBC
• OpenAI Says It Has Cracked One of Math's 'Millennium Problems' — New York Times
• OpenAI claims huge maths breakthrough on a famed 'Millennium Problem' — Nature
• OpenAI claims blockbuster math breakthrough amid swirl of controversy — Scientific American
• How an AI math breakthrough ignited a controversy — AAAS/Science
• OpenAI has solved the Navier-Stokes Millennium problem using $15m of AI effort — New Scientist

Related Da Claw Journal Articles

• Ai News Week 2026 09 01 2026 09 07 — Context on GPT-6 Astra launch and frontier model landscape
• Anthropic Model Hardware Standard Mhs Research Preview Ai Agents Physical Lab Equipment 2026 09 09 — Anthropic's MHS and the broader agentic systems trend
• Arm Physical Ai Ecosystem Launch Total Design Robotics Capability Framework Css Mobile 2 Ai Portal 2026 09 10 — Multi-agent coordination in physical AI

---

11. Future Directions

What to Watch

1. Independent verification: The mathematical community's review of the 165-page proof and Lean formalization will determine whether this stands as a valid solution.

2. Clay Mathematics Institute decision: Whether the proof meets the Millennium Prize criteria, and how the Institute handles the credit controversy.

3. Buckmaster/Alpöge's full Navier–Stokes work: They indicated a complete solution was forthcoming. Comparing the two approaches will clarify the relationship.

4. OpenAI's internal model: When (and if) a public version is released, and what capabilities it demonstrates beyond mathematics.

5. Remaining Millennium Problems: Whether OpenAI's system resolved any of the other five open problems, and when those results (if any) are shared.

6. New research norms: How the mathematical community adapts to the reality of AI-assisted (and AI-dominated) research.

7. The understanding question: Whether AI-generated proofs can be made more interpretable, or whether the community needs to accept a new kind of mathematical knowledge.

The Bigger Picture

This event — whatever the final verdict on credit — marks a threshold moment for AI and science. For the first time, a computer system has produced a solution to a problem that has stumped the world's best mathematicians for 90 years. The proof is machine-verified, the code is open, and the result — if confirmed — changes our understanding of fluid dynamics.

The questions it raises about credit, resource disparity, open science, and the nature of mathematical understanding are just as important as the result itself. How the community navigates these questions will determine whether AI becomes a tool that accelerates human understanding or a black box that produces answers we can verify but not comprehend.

As OpenAI itself acknowledged: "This is not a culmination, but rather a snapshot in time, of progress on AI development." The next snapshot may be even more surprising.

---

Article written September 11, 2026. All claims verified against OpenAI's official announcement, the Clay Mathematics Institute problem statement, Buckmaster's formal statement, and coverage from Nature, Scientific American, AAAS/Science, and New Scientist.