---
title: "OpenAI's Reasoning Model Disproves an 80-Year-Old Math Conjecture Autonomously"
subtitle: "An internal OpenAI reasoning model autonomously cracked the planar unit distance problem, an open question Erdős posed in 1946, using algebraic number theory."
url: "https://alphasignal.ai/news/openai-s-reasoning-model-disproves-an-80-year-old-math-conjecture-autonomously"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "OpenAI"
company_url: "https://alphasignal.ai/company/openai"
published: "2026-06-04T20:19:56.000Z"
updated: "2026-06-04T23:00:14.450Z"
type: "News"
topics:
  - "Reasoning"
  - "Llms"
subtopics:
  - "Math Reasoning"
  - "Test Time Compute"
upvotes: 1463
read_time: "7 min"
source_url: "https://podcasts.apple.com/us/podcast/how-a-reasoning-model-cracked-an-80-year-old/id1820330260?i=1000771190131"
comments: 0
---

# OpenAI's Reasoning Model Disproves an 80-Year-Old Math Conjecture Autonomously

> An internal OpenAI reasoning model autonomously cracked the planar unit distance problem, an open question Erdős posed in 1946, using algebraic number theory.

An internal OpenAI reasoning model has done something that no AI system had pulled off before: it independently disproved a famous open conjecture sitting at the center of an active subfield of mathematics. The result was announced in late May, and OpenAI has now released a [podcast conversation](https://www.youtube.com/watch?v=wNWz5Hbh5VQ) with the three researchers who ran the experiment, Alex Wei, Hongxun Wu, and Lijie Chen, walking through how it actually happened.

The headline is striking on its own. The proof is an important milestone for the math and AI communities, marking the first time that a prominent open problem, central to a subfield of mathematics, has been solved autonomously by AI. But the details of how the model got there are arguably more interesting than the result itself.

## The 80-year-old puzzle

The problem in question is the planar unit distance problem, posed by Paul Erdős in 1946. It is a deceptively simple question that asks how many pairs of points can sit exactly one unit apart on a flat plane, and it has challenged mathematicians since 1946 as one of the best-known questions in combinatorial geometry.

For decades, the working assumption was that you could not do meaningfully better than a square grid. As the number of points grows, grid-like arrangements continue to appear to be remarkably effective. For decades it was widely believed these highly regular structures were about as good as it gets. Erdős himself conjectured that no construction could improve substantially on these intuitive arrangements, even for an extremely large number of points. Formally, he believed the answer was bounded by `n^(1+o(1))`, meaning the number of unit-distance pairs could only grow a hair faster than linearly.

## What the model actually found

OpenAI's model produced a counterexample, not a tight solution. The construction gives a polynomial improvement. The model produced an infinite family of point configurations with at least n^(1+δ) unit-distance pairs for some fixed δ greater than zero.

Princeton mathematician Will Sawin then sharpened the argument shortly afterward. A refinement by Princeton mathematician Will Sawin made the improvement explicit, giving sets of arbitrarily large n points with more than n^1.014 unit-distance pairs. That fixed exponent above 1 is what makes the result decisive: the grid is no longer the ceiling.

The technique is where it gets weird. Instead of bending geometry, the model reached into an entirely different branch of math. The new argument replaces the Gaussian integers by more complicated generalizations from algebraic number theory with richer symmetries that can create many more unit-length differences. The precise argument uses tools such as infinite class field towers and Golod–Shafarevich theory to show the number fields required for the argument actually exist. These ideas were well-known to algebraic number theorists, but it came as a great surprise that these concepts have implications for geometric questions in the Euclidean plane.

In plain terms: the model arranged points by exploiting the symmetries of high-degree number fields. Golod-Shafarevich is a 1964 criterion that guarantees certain infinite towers of number fields exist with the properties you need. Almost no human working on unit distances was looking there.

## One prompt, 125 pages of reasoning

The most surprising operational detail is how little scaffolding was involved. The result is also notable for how it was found. 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. As part of a broader effort to test whether advanced models can contribute to frontier research, we evaluated it on a collection of Erdős problems. In this case, it produced a proof resolving the open problem.

And the prompt itself was open-ended. The model produced a single, very long chain of thought, and the prompt that led to this answer was an open-ended question about whether Erdős's conjecture could be true or false, not an explicit request to prove him wrong. This reasoning is set out in a 125-page document, which the company has not released in full. The model is an experimental, general-purpose reasoning one, not one designed specifically to solve maths problems, and it did all of the work autonomously, in response to a single prompt containing a machine-rewritten statement of Erdős's question.

## How seriously the math community is taking it

The proof was independently scrutinized before being published. The companion paper, [Remarks on the disproof of the unit distance conjecture](https://arxiv.org/html/2605.20695v1), was written by nine mathematicians: Noga Alon, Thomas F. Bloom, W. T. Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood. It presents a short, digested, human-verified version of the recent OpenAI-generated counterexample to the Erdős unit distance conjecture, and a sequence of reflections on it. The argument relies crucially on ideas that may, at least in retrospect, be attributed to Ellenberg-Venkatesh, Golod-Shafarevich, and Hajir-Maire-Ramakrishna.

The endorsements are unusually strong. Tim Gowers, a Fields Medalist, weighed in directly on the official [OpenAI writeup](https://openai.com/index/model-disproves-discrete-geometry-conjecture/): 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. Daniel Litt, who often writes skeptically about AI-math claims, called it "the first result produced autonomously by an AI that I find interesting in itself".

## Why this is different from earlier AI-math hype

OpenAI has been here before, and badly. Seven months ago, the AI giant's former VP Kevin Weil posted on X: "GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others." It turns out, GPT-5 didn't actually solve those problems; it just found solutions that already existed in the literature. Taunts from rivals like Yann LeCun and Google DeepMind CEO Demis Hassabis followed, and Weil promptly took down his premature post.

This time the result is genuinely new mathematics, and the verification pipeline reflects that lesson. Throughout this document, phrases such as "AI proof" or "GPT proof" all refer to the same file, which was first mathematically generated in one shot by an internal model at OpenAI, and then expositionally refined through human interactions with Codex. The math is the model's; the typesetting and exposition were polished collaboratively.

## What it can and cannot do

It is worth being precise about what was achieved. OpenAI's model found a counterexample, not a complete solution. The result disproves Erdős's conjectured n^(1+o(1)) limit by showing that some point sets can do significantly better than the expected grid-style construction. It does not determine the exact maximum possible number of unit-distance pairs. The true maximum remains open. The model showed the grid is beatable, not what the true ceiling is.

The broader picture is also competitive. Days after OpenAI's paper, US mathematician Will Sawin followed the same line of reasoning to an improved result. Also last week, a team from Google DeepMind used one of their own models to resolve nine lesser open problems left by Erdős. Frontier labs are clearly converging on math research as a real benchmark, not a parlor trick.

## What this changes for anyone building with these models

The practical implication is that the same class of model now running coding agents and research assistants can, under the right conditions, produce genuinely novel technical work in narrow domains. A few things stand out:

- The model was not fine-tuned for math, not handed proof tactics, and not given a hint about algebraic number theory. It was a general reasoning system asked an open question.
- The win came from cross-domain transfer, connecting combinatorial geometry to number field towers, an analogy almost no human researcher had pursued.
- Verification still required a small army of human experts. The model produced the argument; nine mathematicians digested, simplified, and signed off on it.

For anyone running long-horizon research workflows, the takeaway is not that AI will replace the mathematician. It is that a general-purpose model, with a sufficiently open prompt and enough compute on its chain of thought, can now surface technical routes that domain experts considered dead ends. Whether that generalizes from a clean problem like unit distances to the messier ones in applied research is the question every lab is now racing to answer.

## Sources

- [X post by @openai](https://x.com/openai/status/2062630454537424930)
- [An OpenAI model has disproved a central conjecture in discrete geometry](https://openai.com/index/model-disproves-discrete-geometry-conjecture/)
- [Remarks on the disproof of the unit distance conjecture](https://arxiv.org/html/2605.20695v1)
- [AI cracks 80-year-old mathematics challenge — researchers are astonished](https://www.nature.com/articles/d41586-026-01651-0)
- [OpenAI claims it solved an 80-year-old math problem — for real this time](https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/)
- [An AI solution to an 80-year-old problem has shocked mathematicians](https://theconversation.com/an-ai-solution-to-an-80-year-old-problem-has-shocked-mathematicians-283686)