OpenAI Just Spent $10M Solving a 90-Year-Old Math Problem — and That Should Terrify You

The Price Tag of Progress

OpenAI just revealed that ten thousand AI agents, churning through roughly three million messages and 130 billion output tokens, solved a partial answer to one of mathematics’ most stubborn problems in 88 hours. At roughly $10 million in compute costs alone, the Navier-Stokes existence and smoothness result represents not just a capability milestone but a new pricing model for what it costs to push against the boundaries of human knowledge.

The problem itself dates back nine decades. The Clay Mathematics Institute attached a $1 million prize to it as one of its seven Millennium Problems. OpenAI clarified immediately it does not intend to claim that prize — or, for that matter, a full solution. The system has addressed two of the four propositions the problem requires. But the gesture alone, spending an estimated $10 million to demonstrate a capability, signals something about where computational mathematics is heading whether the math community likes it or not.

A Race With No Finish Line

What makes this moment sharper than a press release suggests is the timing. NYU professor Tristan Buckmaster and Anthropic researcher Levent Alpoge — both independently working on the same problem — announced their own research trajectories on the same day OpenAI published its results. And then came the complication: both said they had shared research progress with OpenAI through its Codex tool, and they learned afterward that their information had reached the company.

Buckmaster’s statement, released hours after OpenAI’s announcement, accused the company of starting its Navier-Stokes work only after absorbing their research. OpenAI acknowledged the possibility — though calling it low-probability — that de-identified data from Codex users could have contributed to model improvements. The company also insisted its proof process and specific results diverge from those of Buckmaster and Alpoge.

This is the detail that Western coverage has largely flattened. In Korean business and tech media, where mathematical literacy runs deeper among general readers, the story immediately became about intellectual lineage: who invented what, who saw whose work first, and whether the new architecture of AI-assisted research creates conditions for disputes that existing academic norms cannot resolve.

The Korean Lens on a Global Shift

Korean outlets have led international coverage of this story, and the reason is structural. South Korea sits at the intersection of world-class STEM education, a dominant technology sector, and a population acutely aware of how quickly capability can outpace regulation. When a headline declares that an American company solved a 90-year-old problem in three days using machines, the immediate question in Seoul is not wonder but positioning: who controls the next cycle of discovery, and who gets left behind?

That anxiety manifests differently than in Washington or London, where the debate centers on alignment risk or commercial competition. In Korea, the focus lands on institutional displacement — whether universities, research labs, and national science programs can remain relevant when a single corporate entity can redeploy millions of dollars in compute to attack problems that previously consumed decades of collective human effort.

What Actually Changed

The concrete shift is in the economics of mathematical research. For centuries, advancing understanding required human minds working sequentially or in small groups over years. OpenAI’s approach treats the problem as a massive parallel search: generate proofs, test them, iterate, communicate across agents, refine. The system cost $10 million and produced partial results in less than four days. Human researchers, by contrast, spent years on the same problem.

This does not mean human mathematicians are obsolete. The proof OpenAI produced still requires verification by the Clay Mathematics Institute and the broader community. But it does mean the bottleneck is shifting. The constraint is no longer only human creativity or persistence; it is compute capacity and data access — resources concentrated in a handful of companies.

Who Wins, Who Loses

OpenAI wins visibility and a demonstration of capability that no marketing budget could buy. The company’s internal prototype model, trained specifically for rapid mathematical reasoning, now has a proof-of-concept attached to it. Anthropic and other players face pressure to match the infrastructure scale required to compete in this domain.

The losers are harder to name but more consequential. Early-career mathematicians who would have built reputations incrementally through incremental contributions now compete against systems that can ingest and recombine existing research at machine speed. Universities that structure careers around multi-year projects must explain their value proposition when a company can reframe the timeline in days. The Clay Institute, which built its authority on controlling the terms of mathematical verification, now faces a verification pipeline that may need to account for machine-generated proofs it was never designed to evaluate.

What Happens Next

The next six months will clarify whether this is a singular demonstration or a replicable model. OpenAI stated the problem motivated it after hearing rumors that two Millennium Problems had already been solved — meaning the system is reactive as much as proactive. If other labs adopt the same agent-based parallel approach, the pace of partial mathematical breakthroughs could accelerate unpredictably.

Buckmaster and Alpoge’s controversy raises a second-order question about research integrity that will recur. When AI tools trained on user input begin producing results derived, however distantly, from prior human work, the line between assistance and appropriation blurs. Mathematical publishing has no existing framework for attributing contributions that pass through inference engines.

The Korean press is already asking the questions Western outlets have not yet framed: if a corporation can solve 90-year-old problems for $10 million in a week, what happens to the institutions that were built to solve them slowly? The answer to that question will determine whether AI-assisted mathematics becomes a tool for humanity or a mechanism for concentrating the right to define what counts as knowledge.