OpenAI has published a Navier-Stokes proof produced by an internal model and roughly 10,000 collaborating AI agents. The company says the work establishes a finite-time singularity for the famous fluid equations, with an analytical paper and a Lean formalization now available for mathematical scrutiny.

 

The release makes four claims that researchers can now test:

  • A smooth fluid can develop unbounded velocity in finite time.
  • The construction satisfies two official Millennium Problem variants.
  • The agent system reached its result after about 88 hours.
  • GPT-6 Astra completed Lean formalization and verification.

 

OpenAI’s Navier-Stokes Proof Targets a Smooth Forced Blowup

The September 8 research release addresses the existence and smoothness problem for three-dimensional incompressible fluids. Navier-Stokes equations model fluids as continuous media and underpin work on aircraft, weather and blood flow, yet mathematicians have not known whether initially smooth solutions must remain smooth forever.

 

OpenAI’s proposed resolution takes the opposite route. It constructs a vortex that spirals inward and stretches while its speed grows without bound in finite time. The company says the motion begins smoothly at rest, uses a smooth external force and retains finite energy as the singularity forms.

 

That distinction is central to the Millennium Prize formulation. Producing strange behavior with an infinite or badly behaved force would not settle the specified problem. OpenAI says its construction instead establishes statements C and D in the Clay Mathematics Institute’s official formulation, covering smooth forced blowup cases.

 

A mathematical singularity does not mean a real fluid will accelerate to infinite speed. Physical fluids are made of molecules, while the equations use a continuum approximation. A blowup would show that the mathematical description can cease to behave smoothly under conditions that still satisfy its formal assumptions.

 

Quanta Magazine reported that the result has been formally checked in Lean, which increases confidence in its logical validity. The publication also cautioned that mathematicians still need to ensure the encoded theorem precisely matches the intended human problem and assess the proof’s broader significance.

 

Background Reading

 

How 10,000 AI Agents Produced the Navier-Stokes Result

OpenAI says it began training the unreleased internal model on August 28 and found its mathematical performance substantially stronger than GPT-6 Astra. After hearing rumors of progress on Millennium Prize problems, the research team launched coordinated agent groups on several formulations, including variants whose solution would prove regularity and variants that would disprove it.

 

The system first produced a related result for the inviscid Euler equations. Nearly 100 agents worked for about 50 hours on that problem. OpenAI then redirected more resources toward Navier-Stokes, supplied the Euler result to agent groups and used Codex to consolidate promising intermediate ideas across different approaches.

 

The Navier-Stokes effort involved about 10,000 concurrent agents. OpenAI reports that they reached the proposed resolution in roughly 88 hours, sending 2.7 million messages and producing approximately 130 billion output tokens. GPT-6 Astra then spent another 17 hours creating and checking the Lean formalization.

 

Those figures make the result more than a single chatbot answer. It was a large computational search organized around parallel exploration, communication and consolidation. OpenAI has not disclosed the internal model or a complete cost breakdown, so outside researchers cannot yet reproduce the discovery process even though they can inspect the released proof.

 

The distinction between discovery and verification is important. Lean can mechanically check whether each formal step follows from the accepted rules and assumptions. It does not independently decide whether the definitions faithfully capture the original scientific question, whether the proof offers the clearest explanation or how credit should be allocated among prior contributors.

 

Lean Verification Does Not End the Priority Debate

The release arrived amid a dispute involving NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who had announced related work on Euler equations. Both efforts drew on an analytical strategy developed by Diego Córdoba and Luis Martínez-Zoroa for constructing singularities through layered solutions.

 

OpenAI acknowledges that rumors of the other project prompted its effort and credits Buckmaster and Alpöge with priority for their forced Euler result. It says neither its researchers nor agents saw the pair’s work before public release and that no specific user data was accessed, while adding that it cannot exclude indirect model improvement from de-identified product-use data.

 

In a public statement, Buckmaster said he did not know whether the pair’s data had been used and was not making that accusation. His account raises questions about the timeline, communication and recognition of concurrent work, issues that are separate from whether OpenAI’s formal proof is mathematically correct.

 

OpenAI says it will not seek the $1 million Millennium Prize. Formal recognition is not the immediate test in any case: mathematicians will need time to read the analytical argument, compare it with the Lean statement, trace dependencies and determine how the new construction relates to the earlier work that made the approach possible.

 

If the proof survives that review, its importance extends beyond fluid equations. It would show that large agent systems can combine mathematical search, synthesis and formal verification to attack a problem that resisted generations of researchers. The next milestone is not another headline, but sustained independent examination of the public artifacts.