OpenAI Model Disproves Erdős Unit Distance Conjecture (May 2026)

On May 20, 2026, OpenAI announced that one of its general-purpose reasoning models had disproved a central conjecture in discrete geometry that Paul Erdős first posed in 1946, an 80-year-old open problem that generations of specialists had failed to settle.

The conjecture, and why it stood for eighty years

The planar unit-distance problem is one of the oldest open questions in discrete geometry. The statement is deceptively plain. Given n points placed on a flat plane, how many pairs of those points can sit exactly one unit apart? Erdős raised the question in 1946 and went on to conjecture that the maximum number of unit-distance pairs among n points is bounded above by n to the power of one plus little-o of one, written n^(1+o(1)).

For decades the dominant belief among mathematicians was that arrangements based on the square grid were essentially the best possible configurations for maximizing unit-distance pairs. The best known constructions came from carefully chosen subsets of the integer square grid. The best known upper bounds were proved using sophisticated combinatorial geometry. Progress was strictly incremental for nearly eight decades. The field, in effect, had been searching inside a single conceptual neighborhood: the visual plane, the lattice, the compass.

The related Erdős distinct-distances problem was largely settled by Larry Guth and Nets Katz in 2010, which sharpened the contrast. The unit-distance side remained stubbornly open. Erdős believed the truth was very close to the grid. Most working mathematicians believed it with him.

What the OpenAI model actually proved

The new result does not chip away at the conjecture. It breaks it. OpenAI's model exhibited an explicit infinite family of point arrangements that yields at least n^(1+δ) unit-distance pairs for some fixed positive exponent δ greater than zero. Because δ is a fixed positive constant rather than a vanishing quantity, n^(1+δ) grows strictly faster than n^(1+o(1)) as n increases. The lower bound built by the model surpasses the upper bound Erdős had conjectured, and the conjecture falls.

The construction did not arrive by searching harder inside the plane. The model abandoned planar geometry as the working surface entirely. Instead of placing points by hand on a two-dimensional canvas, it encoded points using algebraic numbers whose distances behave predictably under multiplication in algebraic number fields, which are finite extensions of the rational numbers. The familiar Gaussian integers, the complex numbers with integer real and imaginary parts, were the starting frame. The construction then extends those ideas into deeper algebraic number fields, where the distance structure is rich enough to do the counting work.

Two classical tools from algebraic number theory carry the load. The first is infinite class field towers. The second is Golod–Shafarevich theory. In the 1960s, Igor Shafarevich and Evgeny Golod proved that certain class field towers never terminate, and the new proof uses that non-termination directly to manufacture the explicit n^(1+δ) lower bound. Algebraic number theory has rarely served as a primary mechanism in extremal combinatorics on the plane, which is part of why the result reads as a structural surprise as well as a numerical one.

A 125-page chain of thought, written in English

The proof was produced in plain mathematical English, not in a formal proof assistant such as Lean. The model's chain-of-thought, summarized into a readable document and released alongside the announcement, spanned roughly 125 pages and approximately 100,000 tokens of internal reasoning. That artifact is itself part of the story. A 125-page derivation is not a brittle clever trick. It is the kind of sustained, multi-stage construction that working mathematicians spend months writing.

Before the announcement was made public, external research mathematicians checked the argument by hand, line by line. A reviewer assessment circulated alongside the announcement described the work as "a research result that one could recommend acceptance without any hesitation to the Annals of Mathematics." The proof has not yet appeared in a peer-reviewed journal as of May 20, 2026. OpenAI released a write-up and the supporting chain-of-thought document at the same time as the announcement, so that the verification work could continue in the open.

The model used was a general-purpose large language model, not a system specialized for mathematics or geometry. OpenAI did not publicly disclose which specific model was used. The point the company chose to underline was the absence of specialization. An unspecialized system producing this volume of precise logic directly challenges the assumption that solving advanced theoretical mathematics requires bespoke domain-specific programming.

The retraction that set the bar

Claims of AI mathematics breakthroughs are met with skepticism for good reason. In October 2025, OpenAI's then–chief product officer Kevin Weil had announced an earlier claimed mathematical breakthrough that was later retracted. That retraction tightened the standard of evidence the mathematics community would accept the next time the company made a claim of this kind. The May 2026 announcement was constructed with that history in view: the chain-of-thought document was released, external mathematicians were brought in to verify the argument by hand before publication, and OpenAI's framing was narrow.

OpenAI's own description was the line the company chose to defend: "This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics." External experts engaged on the same narrow ground. Sir Timothy Gowers, the Fields Medalist, called the result "a milestone in AI mathematics." Number theorist Arul Shankar said that "AI systems can move beyond assisting mathematicians and begin generating genuinely original ideas." OpenAI researcher Noam Brown's framing of the model was that it was a general-purpose LLM, not targeted at this problem or even at mathematics. The result was also discussed by Will Sawin of Princeton, Noga Alon, Melanie Wood, and Thomas Bloom, who maintains the Erdős Problems online database.

What it changes about how mathematics gets done

The most useful way to read the announcement is methodological. The proof is a single result in one corner of discrete geometry. The pipeline is the durable artifact. A general-purpose reasoning model proposed an unfamiliar bridge, from unit-distance counting in the plane into infinite class field towers and Golod–Shafarevich theory. The model wrote a 125-page argument in plain mathematical English. Human research mathematicians verified that argument line by line before publication. The community is now examining the proof in the open, with the reasoning document available for scrutiny.

That sequence, machine proposes, mathematician verifies, community publishes together, is what the announcement is really putting forward. The result still has to clear peer review. The construction still has to be absorbed by the field and connected to neighboring problems. The 80-year-old question, however, is no longer open. A general-purpose language model with no mathematics specialization produced an infinite family of point arrangements that breaks the n^(1+o(1)) ceiling Erdős set in 1946, and the mathematicians who checked the argument by hand were unable to break it.

Sources

• OpenAI announcement (May 20, 2026)
• Interesting Engineering coverage
• StartupHub.ai write-up
• AutoGPT.net coverage
• Hacker News discussion
• Digg roundup of researcher reactions