An OpenAI reasoning model has disproved a central belief in geometry that stood for nearly 80 years. The problem, first posed by legendary Hungarian mathematician Paul Erdős in 1946, asks a simple question: if you place a set of dots on a flat surface, how many pairs of those dots can sit at exactly the same distance apart? For 80 years, the working belief was that a square grid arrangement was essentially the best anyone could achieve. The model found a completely new class of arrangements that beats the grid, and proved it rigorously.
The proof was not produced by a system trained specifically for mathematics. It came from a general-purpose reasoning model that OpenAI plans to release to users in the near future. That is the detail worth pausing on. Specialized AI tools built for narrow problems have existed for years. A general tool that can independently crack an 80-year-old open problem in geometry, using techniques borrowed from an entirely different branch of mathematics called algebraic number theory, is something different.
The verification is serious. Fields Medalist Tim Gowers, one of the highest honors in mathematics, said he would recommend the proof for publication without any hesitation. Princeton combinatorialist Noga Alon, who described this problem as one of Erdős's personal favorites, called the result an outstanding achievement. The proof was also verified by several other external mathematicians before OpenAI made any public announcement.
This result carries a relevant backstory. In October 2025, OpenAI publicly claimed GPT-5 had solved 10 previously unsolved Erdős problems. That claim collapsed quickly. The model had found existing academic papers containing solutions, not original proofs. Google DeepMind's CEO called it embarrassing, and the posts were deleted. This time, OpenAI published the proof with named reviewers, outside commentary, and specific technical detail. The same mathematician who called out the 2025 misrepresentation is among those now reviewing the work.
For professionals outside mathematics and AI, the practical signal is this: we are watching the boundary shift between what AI can replicate and what it can originate. Until now, AI's value in knowledge work has been speed and coverage: finding existing information faster, summarizing it, drafting from it. A system that can produce verified original reasoning on a problem that stumped human experts for 80 years is pointing toward something qualitatively different.
OpenAI researcher Alex Wei put it plainly: mathematics is a leading indicator of what is coming. The model used algebraic number theory, a branch of mathematics not directly related to the original geometry problem, to crack it. That kind of cross-domain leap, connecting separate fields in a way humans had not explored, is exactly the capability that would matter in applied research: connecting a logistics bottleneck to a supply chain model from a different industry, or linking a chemistry problem to a biological mechanism nobody had paired it with.
The model is not yet publicly available, and the gap between solving an 80-year-old geometry problem and helping a steel manufacturer optimize alloy composition is not trivial. But the direction is clear. The interesting question for business operators is not whether this changes anything today. It is how much of your organization's value sits in the ability to find non-obvious connections across large bodies of knowledge, because that is exactly where these systems are headed next.