OpenAI Claims 90-Year-Old Math Breakthrough With AI Agent Swarm, Igniting Academic Credit Dispute

OpenAI announced on Tuesday that an unreleased internal artificial intelligence model solved the Navier-Stokes existence and smoothness problem, a fluid dynamics puzzle that has stumped mathematicians since French mathematician Jean Leray first formalized foundational questions about it in 1934. The company deployed an autonomous swarm of roughly 10,000 AI agents that worked continuously for 88 hours to produce a mathematical proof, which was subsequently verified using the formal proof assistant Lean over an additional 17 hours^[3].

The achievement targets one of the seven prestigious Millennium Prize Problems established by the Clay Mathematics Institute in 2000, which carries a $1 million bounty for a verified solution. Yet the triumphant technical announcement was immediately clouded by an intellectual property dispute, as outside mathematicians stepped forward to question whether OpenAI leveraged confidential user sessions to steer its automated research swarm toward the breakthrough.

Swarm Intelligence and 130 Billion Output Tokens

According to technical details published by OpenAI, the computational effort was not driven by a single conversational query, but rather by an orchestration of thousands of model instances behaving like an automated research laboratory. The company reported that its internal model is significantly more capable than its previous flagship model, GPT-6 Astra.

To attack the problem, OpenAI partitioned the task across multiple agent groups exploring conflicting mathematical pathways. Teams probed different variations of the problem, with some exploring smooth continuous existence and others seeking counterexamples known as finite-time singularities. In total, the system generated approximately 130 billion output tokens and exchanged 2.7 million autonomous messages on the Navier-Stokes equations alone, while an overarching initiative spanning multiple problems consumed roughly 300 billion tokens^[6].

OpenAI stated that its proof demonstrates that initially smooth three-dimensional fluids can break down under smooth external forcing. The solution outlines a vortex structure that spirals inward and elongates while accelerating rapidly, driving the velocity to infinity in finite time while total energy remains bounded. If upheld, the result establishes that classical continuum equations fail to describe fluid behavior indefinitely under specific conditions^[6].

The Multi-Million Dollar Compute Bill

The sheer scale of compute required to resolve the equations highlights the enormous capital intensity defining frontier AI research. Multiple independent industry analysts noted that the energy and infrastructure costs dwarfed the prize purse offered by the mathematics community.

As reported by Business Insider, LLM benchmark evaluator LisanBench estimated that the 130 billion output tokens alone represented roughly $6.5 million based on average consumer pricing tiers, with total costs potentially reaching $10 million to $40 million when accounting for input token ingestion and context routing. OpenAI stated that running the effort would have cost around $10 million under standard pricing for its most advanced internal models^[9].

OpenAI chief executive Sam Altman joked publicly on social media about the lopsided financial equation, posting on X:

ugh AI is such a bubble, i heard they are selling tokens at a loss, did they know this was only worth $1 million?

Sam Altman, Chief Executive Officer at OpenAI

The company stated that it does not intend to claim the Clay Mathematics Institute bounty.

Academic Collision and the Codex Data Controversy

The announcement triggered immediate pushback from academic circles after New York University mathematics professor Tristan Buckmaster publicly revealed that he and Levent Alpöge, a mathematician who also works at Anthropic, had spent nearly a year pursuing a closely related proof. The researchers had used OpenAI tools, specifically Codex, as a computational workspace for their drafts and formulations.^[4]

According to reporting from VentureBeat and TechRepublic, Buckmaster contacted OpenAI on September 3 after learning rumors were circulating about mathematical milestones, seeking clarification on whether OpenAI was targeting the same specific smooth-forcing framework his team had privately developed. Buckmaster stated that OpenAI launched its swarm push on September 1 precisely because of chatter surrounding his upcoming work.

A contentious exchange ensued. Buckmaster alleged that OpenAI researcher Sébastien Bubeck initially offered him sole credit on a joint publication if Alpöge's name was removed due to his employment at rival Anthropic, an accusation Bubeck disputed on social media. OpenAI defended its conduct by asserting that researchers did not manually inspect or look up user account sessions. However, in an official statement provided to VentureBeat, the company conceded an important qualification, acknowledging it cannot rule out that de-identified data derived from usage of its products helped train and improve the models that subsequently solved the problem.

Writing in MIT Technology Review, journalist Grace Huckins observed that the data question exposes an unresolved debate regarding scientific taste. E^[11]ven if the AI generated the formal arguments, human intuition and deliberate problem formulation by specialized researchers appeared indispensable in steering the models toward the productive pathway.

Verification Bottlenecks in Pure Mathematics

Despite OpenAI publishing its proof along with machine-checked code in Lean, the mathematical community remains cautious. A^[8]cceptance of a Millennium Prize Problem requires an exhaustive vetting protocol that computational verification cannot instantly circumvent.

As Clay Mathematics Institute president Martin Bridson told AFP, the institute's evaluation protocols remain deliberately measured and require work to withstand community examination over multiple years before a committee convenes to review the result. While machine verification reduces arithmetic slips, mathematicians must still inspect whether the formal definitions precisely align with the historical conditions Jean Leray set down nine decades ago.