1. 首页
  2. ›OpenAI Deploys 10,000+ AI Agents to Crack Millennium Math Problem in 88 Hours

OpenAI Deploys 10,000+ AI Agents to Crack Millennium Math Problem in 88 Hours

N
nashnova research
·2026-11-15发布·阅读约 13 分钟

OpenAI marshalled roughly 10,000 AI agents over 88 hours to prove the Navier–Stokes existence-and-smoothness problem — a million-dollar challenge unsolved for over 90 years — signaling that AI math capability is leaping from olympiad medals to frontier research, and repricing the return calculus on compute investment.

01

How hard is this problem, and how did AI solve it?

•
The Navier–Stokes equations — the fundamental math describing fluid motion in everything from aircraft design to weather forecasting to blood flow — are one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000, carrying a $1 million bounty. Human mathematicians spent over 90 years without a solution.
•
The core finding: a fluid starting from rest, driven by a smooth external force, can spontaneously develop a "singularity" in finite time. In plain terms = imagine a strand of spaghetti being twisted ever thinner and spinning ever faster until it mathematically "blows up" — the fluid generates this extreme state on its own.
•
The approach was brute-force collaboration: roughly 10,000 AI agents worked in groups, sharing information within teams, equipped with cached internet access and code-execution environments. They exchanged about 2.7 million messages and generated roughly 130 billion output tokens over approximately 88 hours.
•
After the proof was complete, GPT-6 Astra spent another ~17 hours line-checking the entire argument using Lean — a formal verification tool that mechanically confirms each logical step.
02

What did it cost, and how fast will costs fall?

•
Sam Altman estimated the compute cost at roughly $10 million. OpenAI said it does not intend to claim the $1 million prize.
•
Altman made a bet: within one year of the Navier–Stokes proof, repeating an equivalent task will cost one hundredth as much — around $100,000.
•
This means → if the bet holds, a $10-million-class AI research task today becomes a mid-range R&D budget item in twelve months. The speed of compute-cost decline directly determines how economically viable it is for AI to tackle "major science."
03

Anthropic solved one too — who was first?

•
The trigger was a rumor: on September 1, OpenAI heard that two Millennium Prize Problems had already been solved, and launched its own effort.
•
The truth: Anthropic employee Levent Alpöge and NYU math professor Tristan Buckmaster, using an internal Anthropic model, proved the forced Euler equations — related to but distinct from Navier–Stokes.
•
OpenAI acknowledged in its announcement that Anthropic's team completed the forced-Euler proof first, though the two proofs differ significantly in method. This reflects a new stage: the math race between top AI labs has escalated from "who wins the olympiad first" to "who cracks a Millennium Problem first."
04

What is happening with the Yang–Mills problem?

•
In late October, OpenAI announced it had completed the first half of a proof for the Yang–Mills existence problem — another Millennium Prize Problem tied to the mathematical foundations of the Standard Model in particle physics. The full proof requires both "existence" and "mass gap"; the harder second part remains unsolved, and the $1 million prize is still open.
•
To ensure credibility, OpenAI invited David Gross, 2004 Nobel laureate in physics, to chair a five-person "escort committee" for pre-publication review. All Lean code and review documents were released on GitHub.
•
The review document opens with: "We used OpenAI's Astra to help digest the submitted Lean code. We cannot yet claim to fully understand this proof." In plain terms = even a Nobel-led review panel needed AI assistance to read AI-written proofs — and the acknowledgments note that Anthropic's Fable model was also consulted during revisions.
05

What does this mean for the AI industry and markets?

•
Podcast host Dwarkesh Patel laid out a timeline: in 2023, models couldn't reliably add four-digit numbers; in 2025, they won International Math Olympiad gold; in 2026, they solved a Millennium Problem.
•
He revised his forecast accordingly: AI matching top human researchers arrives "by 2030 at the latest — possibly next year."
•
This means → the pace of AI's problem-solving leap is repricing market expectations for AI R&D cycles and compute-investment returns. If Altman's "costs drop 100× in a year" call proves roughly right, the payback period on today's massive compute-infrastructure bets may be shorter than the market assumes.

市场有风险,内容仅供研究参考,不构成投资建议。

相关标的

4
  • MicrosoftMSFT.US+0.2%

    As OpenAI's largest investor and compute provider, tens of millions of dollars in compute consumption translate directly into Azure revenue.

  • NVIDIANVDA.US-2%

    Parallel inference across 10,000+ agents relies on massive GPU clusters; the surge in compute demand directly benefits AI chip suppliers.

  • Robotics & Artificial Intelligence ETFBOTZ.ETF-0.6%

    The leap in AI capabilities is driving a reassessment of R&D timelines and commercialization expectations across the entire AI value chain, and this ETF covers core AI holdings.

  • Arista NetworksANET.US-4.5%

    Exchanging 2.7 million messages across 10,000+ agents requires high-bandwidth data center networking; Arista is a leading AI cluster network supplier.