On 20 May 2026, OpenAI announced that one of its general-purpose reasoning models had disproved a major unsolved conjecture in discrete geometry: the planar unit distance problem first posed by Paul Erdős in 1946. The model found point arrangements that produce significantly more unit-distance pairs than the square-grid layouts mathematicians have used as the working ceiling for nearly 80 years.
The problem sounds elementary. Place n dots on a flat plane. How many pairs of those dots can sit exactly one unit apart? For decades, the best-known constructions used variations of the square grid, and the working assumption was that you could not do much better. OpenAI's model, by their own account, found an infinite family of arrangements that do.
The proof did not come from a maths specialist
The detail that has mathematicians talking is what kind of model produced the result. According to OpenAI's own writeup, the proof was generated by a general-purpose reasoning model, not by a system trained specifically for mathematics, scaffolded to search through proof strategies, or tuned for the unit distance problem in particular. Instead of relying on the geometric tricks that the field has used for decades, the model connected the problem to algebraic number theory, a deep and somewhat unrelated area of mathematics.
OpenAI published companion remarks from Noga Alon, Melanie Wood, and Thomas Bloom, all respected mathematicians, who independently verified the proof.
What it does and does not mean
This is not the first time a language model has produced new mathematics. DeepMind's AlphaGeometry and AlphaProof, and earlier collaborations between mathematicians and large models, have all chipped at open problems. TechCrunch's coverage notes, sensibly, that the result is the latest in a sequence rather than a single moment of arrival.
What is genuinely new is the absence of scaffolding. The model was not handed the problem as a search task. It identified a connection between two areas of mathematics and used it to construct a counterexample. That is the kind of move you would expect from a doctoral student in their second PhD year, not a chatbot.
The natural caveats apply. The unit distance problem is a famous open question but it is not load-bearing for any field of applied mathematics. The proof has been verified, but the broader claim that this is a general capability rather than a one-off result still depends on further examples. AI Claudius covered Jack Clark's Oxford prediction of Nobel-grade AI research within a year. This result will be cited in support of that prediction. It does not settle the question.
Commentarii · 0