Skip to main content
Back

OpenAI's Reasoning Model Disproves an 80-Year-Old Erdős Conjecture

An OpenAI reasoning model cracked an 80-year-old geometry conjecture using algebraic number theory — the first time AI has autonomously solved a prominent open math problem.

  • First autonomous AI proof of a major open math problem: OpenAI's internal reasoning model disproved the Erdős unit distance conjecture, unsolved since 1946.
  • Key result: The model found point configurations with n^1.014 unit-distance pairs, beating the square-grid ceiling mathematicians assumed for 80 years.
  • Surprising method: The model applied algebraic number theory (infinite class field towers, Golod-Shafarevich theory) to a geometry problem — a connection no human had pursued.
  • Validated by top mathematicians: Fields Medalist Tim Gowers said he would recommend it for the Annals of Mathematics "without any hesitation."
  • Competitive context: Google DeepMind solved nine other Erdős problems the next day; GPT-5.5 can also reproduce the proof with a small hint.
  • Open problem remains: The gap between the new lower bound (n^1.014) and the best upper bound (n^1.333) is still unresolved — more work needed.

For 80 years, one of the most famous unsolved problems in combinatorial geometry sat untouched. Mathematicians knew roughly how many pairs of points could sit exactly one unit apart in a flat plane, and most believed Paul Erdős had essentially nailed the ceiling back in 1946. Then an OpenAI reasoning model proved him wrong.

The planar unit distance problem, first posed by Paul Erdős in 1946, is one of the best-known questions in combinatorial geometry: easy to state and remarkably difficult to resolve. The question is deceptively simple: place n points in a 2D plane. How many pairs of those points can be exactly distance 1 apart? Since Erdős's original work, the prevailing belief was that square grid constructions were essentially optimal for maximizing the number of unit-distance pairs. An internal OpenAI model has now disproved this longstanding conjecture, providing an infinite family of examples that yield a polynomial improvement.

What Erdős actually conjectured

Erdős's conjecture was that the actual optimum was much closer to the lower bound than the upper one. He predicted, but couldn't prove, that the maximum number of unit-distance pairs grows just barely faster than the number of points. In math notation, he conjectured the maximum was n^(1+o(1)), where the extra term in the exponent shrinks toward zero as n grows. The best upper bound, O(n^(4/3)), dates to 1984 and had not meaningfully improved since.

The OpenAI model disproved this longstanding conjecture, discovering an infinite family of constructions using deep algebraic number theory , specifically Golod-Shafarevich theory and infinite class field towers , that achieve a polynomial improvement over square grids, specifically n^(1+δ) unit-distance pairs for some fixed δ > 0. The improvement has been quantified with an exponent of approximately 0.014, a figure later refined by Princeton mathematician Will Sawin.

A bridge from number theory to geometry

The most striking thing about the proof is not just that it worked, but how. Rather than iterating on known grid arrangements, the model approached the problem through algebraic number theory, connecting it to advanced mathematical structures called infinite class field towers. The result is an infinite family of configurations that surpass the traditionally accepted optimal ones, refuting Erdős's conjectured upper bound outright.

To understand the leap: Erdős's original construction used the Gaussian integers , numbers of the form a + bi where a and b are regular integers and i is the square root of -1. These form a well-understood number system with clean factorization properties. The AI's proof replaces these with far more complex generalizations from algebraic number theory , extensions of the integers with richer symmetries that can pack in many more unit-length differences. Bloom's commentary credits the model with succeeding precisely because it did not share the human bias of presumed truth, and instead systematically applied techniques from algebraic number theory.

How the model actually worked

The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular. It was a proof produced during a set of Erdős problem evaluations , essentially a stress test of the model's frontier reasoning capabilities.

The model's chain of thought ran to 125 pages before outputting the final proof. Reading through it, it doesn't resemble a human having an epiphany , it reads more like a machine executing a depth-first search through the mathematical literature, trying polynomials, finite fields, elliptic curves, Dirichlet units, and graphs, running down every branch until it hits a contradiction, then backtracking. The model reaches the crucial idea , which it describes in its own reasoning as "frightening" , on page 39.

OpenAI also published data on success rates at different compute budgets. Even at maximum token budget, the internal model only solves the problem about half the time. That's a meaningful signal: this wasn't a one-shot miracle, it was a probabilistic outcome from a powerful reasoning process run at scale.

What the mathematicians said

OpenAI brought in external validators before the announcement, and the responses were unambiguous:

  • Tim Gowers, who won the Fields Medal , the most prestigious prize in mathematics , wrote that "there is no doubt that the solution to the unit-distance problem is a milestone in AI mathematics."
  • Number theorist Arul Shankar stated 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."
  • University of Toronto professor Daniel Litt wrote that "this is the first example of a result produced autonomously by an AI that I find exciting in itself, as opposed to as a leading indicator."
  • Gil Kalai compared it to the computer-assisted proof of the Four Color Theorem by Appel and Haken in 1976, suggesting its significance extends beyond combinatorics.

Gowers also admitted that when he first heard the result, he assumed the AI had proved the conjecture , and spent an evening recalibrating his worldview. When he learned it had disproved it instead, he described it as "a big relief." He then spent the rest of his commentary exploring whether that relief was actually justified.

The credibility gap OpenAI had to close

This result lands against a specific backdrop of prior overclaiming. In October 2025, OpenAI's then-VP Kevin Weil posted that GPT-5 had "found solutions to 10 previously unsolved Erdős problems." The claim quickly fell apart: Thomas Bloom, the mathematician who maintains the erdosproblems.com database, called it "a dramatic misrepresentation." GPT-5 had not solved anything new , it had simply surfaced existing published solutions. Meta's Yann LeCun mocked the claim publicly, and Google DeepMind CEO Demis Hassabis called it "embarrassing."

Seven months later, the same mathematicians who debunked that earlier claim are now endorsing this one. Thomas Bloom is among the co-authors of the companion paper. That reversal is arguably as significant as the math itself.

The competitive landscape just shifted

Days after OpenAI's paper, Princeton mathematician Will Sawin followed the same line of reasoning to an improved result. Also that same week, a team from Google DeepMind used one of their own models to resolve nine lesser open problems left by Erdős. The two systems use fundamentally different approaches: OpenAI relies on natural language reasoning with human verification, while Google combines language models with machine-checkable proofs.

Shortly after OpenAI's announcement, University of Michigan postdoc Xiao Ma found that GPT-5.5 , a publicly available model , was also able to prove Erdős wrong if given a small hint. That raises an uncomfortable question: if a consumer-grade model can crack an 80-year-old conjecture with a nudge, how many other open problems are already within reach of models people use every day?

What this does and doesn't resolve

It's worth being precise about the limits of the result:

  • The AI proof establishes that u(n) grows at least as fast as n^1.014 for infinitely many values of n.
  • The result doesn't completely resolve the problem. The best upper bound for the number of unit distances is still around n^1.333. More work is needed to close this gap.
  • The proof spans roughly 125 pages. Human mathematicians then rewrote and extended it.
  • The model did not pioneer new mathematical techniques , it applied existing tools from algebraic number theory in a context no one had thought to try.

What assumptions need updating

The deeper shift here is about what AI reasoning systems are actually capable of. The headline is not that AI is now smarter than mathematicians, but that pure chain-of-thought reasoning , without elaborate symbolic scaffolding , accomplished a result that previously required heavy computer-aided math infrastructure.

Two properties of the solution played directly to AI's structural advantages over humans. First, the proof required connecting a geometry problem to algebraic number theory , a combination that almost no individual mathematician would hold simultaneously. Second, the reasoning process was such a grind , and seemingly unlikely to succeed , that most humans would not have thought it worth the trouble. An AI has no such reluctance.

The result also opens a new mathematical frontier. As Thomas Bloom writes in the companion paper, the unexpected connection between algebraic number theory and discrete geometry "shows that there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected." Algebraic number theorists are now expected to take a close look at other open problems in discrete geometry , a research direction that simply didn't exist before this proof.

The full proof, the companion remarks from external mathematicians, and an abridged version of the model's chain of thought are all publicly available. For anyone who wants to understand what frontier AI reasoning actually looks like in practice, the chain of thought document is worth reading , not because it reads like genius, but because it reads like something genuinely new.