An artificial intelligence model built by OpenAI has disproved a famous conjecture in discrete geometry that had stood unsolved for nearly eight decades, a result mathematicians are calling a landmark for machine reasoning. OpenAI announced the breakthrough on May 20, 2026, centered on the planar unit distance problem first posed by the Hungarian mathematician Paul Erdos in 1946.
An 80-year-old problem falls
The unit distance problem asks a deceptively simple question: if you place a number of points on a flat plane, how many pairs of them can sit exactly one unit apart? For generations, researchers believed that square-grid arrangements were essentially optimal, and Erdos proposed that the count of unit-distance pairs could grow only slightly faster than linearly as more points were added.
According to OpenAI, its model overturned that long-held assumption by discovering an infinite family of point arrangements that produce significantly more unit-distance pairs than the classic grid. Princeton mathematician Will Sawin later refined the work, showing the improvement could be captured by a fixed exponent rather than a marginal gain.
Number theory, not geometry tricks
What surprised experts most was the method. Instead of leaning on conventional geometric techniques, the model connected the puzzle to algebraic number theory, drawing on advanced tools such as infinite class field towers and Golod-Shafarevich theory that are rarely associated with this kind of geometry. In effect, the system exploited hidden symmetries inside exotic number systems to pack in many more one-unit distances.
OpenAI emphasized that the proof came from a general-purpose reasoning model rather than specialized theorem-proving software, and that engineers did not train it on the unit distance problem or build dedicated search tools for the task. The model received the problem statement and produced the solution on its own.
Mathematicians weigh in
The argument was checked by a group of external mathematicians who produced a companion paper explaining its significance. Fields Medal winner Tim Gowers described the achievement as "a milestone in AI mathematics," while number theorist Arul Shankar said the work shows AI systems can move beyond assisting researchers to generating genuinely original ideas. Thomas Bloom, who contributed to the companion work, suggested deep number theory may hold answers to other open questions in discrete geometry.
Why it matters
The result lands amid a rapid acceleration in AI reasoning capabilities. Just months earlier, the company behind the model completed a record fundraising effort, as detailed in our report on how OpenAI raised a record sum in a historic AI funding round. The pace of frontier model releases has been relentless, from new systems built to give machines muscle memory for delicate manual tasks to rival efforts such as Google's Gemini 3.5 Flash unveiled at I/O 2026. Researchers say models that can sustain long chains of reasoning could eventually contribute to physics, biology, engineering and medicine. For now, a problem that resisted human effort for almost 80 years has fallen to a machine that approached it from an entirely unexpected direction.
Reporting based on coverage from OpenAI and Interesting Engineering.