OpenAI’s General Model Cracks 80-Year-Old Math Problem, Overturning Geometry Conjecture
On May 21, 2026, OpenAI released a research paper announcing that one of its general reasoning models successfully solved the famous unit distance problem, an open question that has stumped the mathematical community for nearly 80 years. The model not only autonomously provided a complete proof but also surprised mathematicians with its unconventional methods, marking a milestone step for AI’s participation in frontier scientific exploration.
1. The Unit Distance Problem: A Nearly 80-Year-Old Puzzle
The unit distance problem was posed by Hungarian mathematician Paul Erdős in 1946 and is one of the central problems in combinatorial geometry. The question is: given n points in a plane, what is the maximum number of pairs of points that are exactly one unit apart? This maximum value is denoted by the function u(n).
Over the past 80 years, the best-known construction (the lower bound) found by mathematicians comes from a scaled square grid, where the number of pairs grows at a rate of approximately n1 + c/log log n, which is only slightly faster than linear growth. Based on this, it was widely conjectured that the upper bound for u(n) should be n^(1+o(1)), where the o(1) term tends to 0 as n increases. Erdős himself proposed a related conjecture. However, this long-held belief has now been overturned by the AI’s discovery.
2. The AI’s New Construction and Improved Lower Bound
OpenAI’s model constructed a new family of point sets, proving that for infinitely many values of n, the lower bound for the number of unit distance pairs is at least n1+ε, where ε is a fixed positive exponent. This fundamentally shows that the number of unit distances can grow significantly faster than any previously known construction.
Although the AI’s original proof did not provide a specific value for ε, Professor Will Sawin of the Princeton University Department of Mathematics, in a subsequent analysis, pinpointed the exponent as ε ≈ 1/log(12) ≈ 0.402. This result represents the first major breakthrough on the lower bound since 1946, breaking a long-standing stalemate. The proof’s validity has been independently verified by a group of external mathematicians, who also wrote an accompanying paper explaining the background and significance of the achievement.
3. An Interdisciplinary Proof: The Surprising Application of Algebraic Number Theory
The most striking aspect of this breakthrough is not just the result itself, but the path taken to prove it. The AI’s proof did not follow traditional geometric or combinatorial methods but instead unexpectedly introduced profound tools from algebraic number theory.
Specifically, the core idea of the proof generalizes the Gaussian integers (complex numbers of the form a+bi) used in Erdős’s original construction to algebraic number fields with richer symmetries. To ensure the existence of the required number fields, the proof utilizes advanced concepts from algebraic number theory such as infinite class field towers and the Golod-Shafarevich theory. This cross-disciplinary connection, applying abstract tools from pure mathematics to solve a geometric problem in the Euclidean plane, is unprecedented.
Fields Medalist Timothy Gowers described it as “a milestone for AI in mathematics.” Princeton number theorist Arul Shankar also stated that this demonstrates AI’s capability to generate truly original and sophisticated mathematical ideas.
4. General Models and the Future of Scientific Research
The feat was accomplished not by a specialized system designed for a specific mathematical task, but by one of OpenAI’s general reasoning models. According to OpenAI, the model solved the problem without targeted training, as part of an evaluation of its frontier research capabilities. This is the first time an AI has autonomously solved a famous open problem at the heart of a mathematical subfield.
The significance of this event transcends the solution to a single problem. It showcases the powerful potential of advanced AI models in handling complex reasoning, making interdisciplinary connections, and exploring paths unenvisioned by human researchers. OpenAI believes this capability can be applied to many other fields, such as biology, physics, and materials science, thereby accelerating the process of scientific discovery. In the future, AI is poised to become a powerful collaborator for scientists, while human expertise and judgment will become even more crucial in setting research directions and interpreting results.
