An OpenAI Model Disproves an 80-Year-Old Geometry Conjecture

OpenAI said on Wednesday that one of its reasoning models had disproved a long-standing conjecture in discrete geometry, the first time an AI system has autonomously solved a prominent open problem in mathematics. The claim, published by the company on its research site, has been backed by independent mathematicians who reviewed the proof.

The problem traces back to Paul Erdos, who posed it in 1946. Known as the unit-distance problem, it asks how many pairs of points in a flat plane can sit exactly one unit apart. For nearly eighty years, mathematicians believed the most efficient configurations resembled square grids. The OpenAI model overturned that assumption.

How the result came about

According to OpenAI, the proof was produced by a general-purpose reasoning model rather than a specialised theorem prover. Engineers did not train it on the unit-distance problem specifically, nor did they build dedicated search tools for the task. The model reportedly reached for deep number theory to construct its counterexample, an approach Interesting Engineering noted was not anticipated by researchers in the field.

TechCrunch reports that the verification step matters more than the result itself. OpenAI made a public claim last year about cracking other Erdos problems, only to retract after independent mathematicians showed the work was flawed. This time the same critics have signed off.

What it means for the field

If an off-the-shelf reasoning model can knock over a problem that resisted eighty years of human attention, the question of what else might fall is suddenly live. Mathematicians have used software as a research aid for decades, but autonomous discovery without bespoke training is new ground. Whether this generalises to other branches of mathematics or remains a one-off success will depend on what OpenAI's model, and its competitors, can show next.

The result also fits a wider pattern at the frontier labs. Anthropic, Google DeepMind and OpenAI have all pointed to reasoning gains as the main lever for the next year of progress, after pre-training gains began to plateau. A disproved conjecture is concrete evidence that something is moving.