An OpenAI model refuted on May 20, 2026, an 80-year-old conjecture in discrete geometry, marking a historic first for artificial intelligence in fundamental mathematical research.artificial intelligence in fundamental mathematical research.
A problem posed in 1946, never solved since
The planar unit distance problem (or planar unit distance problem) was formulated by Hungarian mathematician Paul Erdős in 1946. The question is simple to state: if we place n points on a plane, what is the maximum number of pairs of points separated by a distance exactly equal to 1? For nearly 80 years, the mathematical community estimated that square grids constituted the best possible arrangement. Erdős himself considered this problem one of his favorites and had even offered a financial reward to anyone who solved it.
The best known construction, based on Gaussian integers, produces approximately n^(1+C/log log n) unit distance pairs, a progression barely greater than linear. This result had stood for decades, resisting the attempts of hundreds of mathematicians.
The unexpected method of OpenAI's model
According to OpenAI, the model discovered an infinite family of point arrangements producing a significantly higher number of unit distance pairs than the classic square grid approach. Princeton mathematician Will Sawin then refined the result and showed that the improvement could be expressed with a fixed exponent. This exponent was specified as δ = 0.014 by Sawin.
What makes the result particularly remarkable is the mathematical tool employed: infinite class field towers and Golod-Shafarevich theory, from algebraic number theory, applied to a problem in fundamental Euclidean geometry. This connection between two seemingly distant fields is, according to the mathematicians involved, the core of the result's originality.
The model did not use brute force. It constructed a 125-page reasoning by linking two branches of mathematics that human researchers had not thought to associate for this problem, which is precisely the kind of creative process previously attributed to human mathematical intuition.
Independent validation that changes everything
OpenAI learned from an embarrassing episode that occurred in October 2025. At the time, Vice President Kevin Weil claimed on X that GPT-5 had solved ten still-open Erdős problems. The claim quickly collapsed: Thomas Bloom, the mathematician who maintains the site Erdos Problems, showed that the model had produced no original proof, but had simply found solutions already published in the existing literature. Bloom described it as "a dramatic misrepresentation" (a dramatic misrepresentation).
This time, OpenAI published an accompanying document signed by several leading mathematicians who independently verified the proof. Fields Medalist Tim Gowers reviewed the work, as did Princeton's Will Sawin, and both validated the correctness of the proof. Noga Alon, a combinatorialist at Princeton who knew Erdos personally, and Thomas Bloom himself, the very person who had exposed OpenAI's previous failure, also provided favorable statements.
Tim Gowers stated in the accompanying document: "There is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics." (There is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics: if a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation.)
A generalist model, not a system dedicated to mathematics
OpenAI specified that this result comes from a new general-purpose reasoning model, not from a system designed specifically to solve mathematical problems or this problem in particular. It is this distinction that captures researchers' attention. If a specialized mathematics model achieves a mathematical breakthrough, the result is impressive but expected. If a general-purpose reasoning model, without specific configuration or external help, achieves it autonomously, it suggests something much broader about the reasoning capabilities of current systems.
According to Arul Shankar, a renowned number theorist: "This proof shows that current AI models go beyond the simple role of assistants for human mathematicians: they are capable of having original and ingenious ideas, and then carrying them out." (In my opinion this paper demonstrates that current AI models go beyond just helpers to human mathematicians – they are capable of having original ingenious ideas, and then carrying them out to fruition.)
Implications that go beyond pure mathematics
OpenAI placed this advance within the framework of a more ambitious approach to automated research, believing that similar capabilities could one day support work in biology, physics, materials science, and medicine, where many problems are too vast or too complex to be handled by traditional teams.
Noam Brown, a researcher at OpenAI, wrote on X: "Less than a year ago, state-of-the-art AI models were achieving gold medal level at the International Mathematical Olympiad. I expect this pace of progress to continue." (Less than 1 year ago frontier AI models were at IMO gold-level performance. I expect this pace of progress to continue.)
This announcement comes in the same week that Jack Clark, co-founder of Anthropic, stated at a conference in Oxford that AI would contribute to a Nobel Prize-winning discovery within twelve months. OpenAI's breakthrough in discrete geometry is not an isolated event: it is part of a visible acceleration of the capabilities of systems AI reasoning on open scientific problems. For researchers and mathematicians, the question is no longer whether AI can contribute to fundamental research, but how quickly and how far these systems will go.
No Comment! Be the first one.