2026-05-21 09:36 UTCIn-site rewrite2 min readUpdated: 2026-06-30 13:03 UTC

Fields Medalist Stunned: OpenAI's Non-Mathematical Model Independently Solves 80-Year-Old Unsolved Math Problem

An internal general-purpose reasoning model from OpenAI has cracked the Erdős unit distance problem, a classic challenge that has seen no substantial progress for 80 years. Fields Medalist Timothy Gowers called it a clear case of an AI autonomously achieving a mathematical breakthrough. The solution spans 125 pages of reasoning.

Source量子位Author: 闻乐

Article intelligence

EngineersAdvanced

Key points

The model discovered a novel algebraic number theory construction, proving super-linear growth for unit distances, overturning the long-held belief of linear growth.
Gowers stated this is the first clear instance of an AI independently solving a famous unsolved math problem.
Unlike previous AI claims that were debunked, this achievement has been positively received by the mathematical community.

Why it matters

This matters because the model discovered a novel algebraic number theory construction, proving super-linear growth for unit distances, overturning the long-held belief of linear growth.

Technical impact

May affect model selection, inference cost, product capability, and evaluation benchmarks.

This panel is AI-generated and reviewed for accuracy.

OpenAI has made a significant breakthrough in mathematics with a general-purpose reasoning model, not a specialized one, that has solved a classic problem posed by the legendary mathematician Paul Erdős in 1946. The Erdős unit distance problem asks: given n points on a plane, what is the maximum number of pairs that can be exactly one unit apart? Despite its deceptive simplicity, the problem had resisted progress for nearly 80 years, with mathematicians converging on the belief that the number of unit-distance pairs grows roughly linearly with n.

OpenAI's internal model, described as a general reasoning AI, took a different approach. Instead of geometric methods, it employed algebraic number theory to construct a novel arrangement of points. The result was a proof that the growth rate is actually super-linear: u(n) ≥ n^(1+δ) for some δ > 0, contradicting the long-held consensus that the exponent approaches 1. In other words, the previously assumed "near-zero" correction was actually positive.

Fields Medalist Timothy Gowers, a renowned mathematician, publicly praised the achievement, calling it "the first clear case of an AI solving a very famous unsolved mathematical problem" and "a genuine mathematical breakthrough by an AI." He emphasized that this was not a case of the AI finding a known result but an original discovery. The OpenAI model's reasoning, even after simplification, filled 125 pages, with one section describing its construction as "frightening."

The breakthrough is especially notable because it comes from a general-purpose AI, not a specialized mathematics model. Noam Brown, an OpenAI researcher, indicated that the model will be released soon, raising excitement about what other hidden capabilities the AI possesses.

This achievement marks a contrast to previous AI hype in mathematics. In October 2025, OpenAI's VP Kevin Weil claimed that GPT-5 had solved ten Erdős problems, but mathematician Thomas Bloom quickly debunked this, showing that GPT-5 had merely retrieved existing papers unknown to Bloom. DeepMind's Demis Hassabis called the claim "embarrassing." Now, seven months later, Bloom offered a very different assessment: "This is the most impressive achievement of AI in mathematics so far."

The result is a major milestone in AI-driven mathematical research, demonstrating that general-purpose models can make original contributions to deep mathematical problems. The full implications for the future of mathematical discovery remain to be seen, but the 125-page proof may inspire new lines of inquiry and further AI-assisted breakthroughs.