OpenAI Just Settled an 80-Year-Old Erdős Conjecture: Inside the Unit-Distance Proof

7 days ago
3 min read

On May 20, 2026, OpenAI published a short note and a companion mathematical writeup claiming that one of its general-purpose reasoning models had autonomously disproved a long-standing conjecture in discrete geometry: the planar unit-distance problem, originally posed by Paul Erdős in 1946. If the result withstands broader mathematical scrutiny, this may represent one of the clearest examples yet of an AI system producing a novel proof for a major open mathematical problem without a publicly disclosed human-supplied proof sketch.

We want to be careful here because OpenAI made a similar Erdős-related claim in October 2025 that drew sharp criticism. So below we walk through what the problem actually says, what was proved, who is independently endorsing the result, and what builders should take away from it.

The Question

The unit-distance problem asks a deceptively simple thing. Place n points in the plane. How many pairs of those points can be exactly distance 1 apart?

For 80 years the working intuition was that the square integer grid was essentially the optimal arrangement. Erdős himself proposed that the count should grow only slightly faster than linearly in n, with a tiny correction term tied to the divisor structure of integers. That intuition shaped a sub-area of combinatorial geometry that has produced hundreds of papers.

What the Model Actually Proved

According to OpenAI's accompanying mathematical writeup, the reasoning model constructed an infinite family of point arrangements that produces meaningfully more unit-distance pairs than standard square-grid constructions at the same n. The proof imports machinery from algebraic number theory, in particular the structure of certain algebraic integers, to define point sets whose pairwise distances cluster on a richer set of common values than a lattice can support.

The result is published as a set of mathematical remarks at the OpenAI CDN. The PDF link in the press materials is cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-remarks.pdf. The exact reasoning model variant used was not disclosed in the announcement.

The Independent Reviewers Matter

Three working mathematicians provided accompanying commentary: Noga Alon (Princeton, Tel Aviv), Melanie Wood (Harvard), and Thomas Bloom (Oxford). Bloom in particular is worth noting. He publicly characterized OpenAI's October 2025 Erdős claim as a dramatic misrepresentation. This time he wrote that AI is helping the field more fully explore the cathedral of mathematics and asked what other unseen wonders are waiting in the wings. That shift in tone from a vocal prior critic is one of the strongest public signals so far that the result is being taken seriously by working mathematicians.

Why This Is Different From the 2025 Misfire

The October 2025 episode involved OpenAI publicizing alleged solutions to Erdős problems that turned out, on inspection, to be restatements of known results or invalid arguments. Critics including Yann LeCun and Demis Hassabis pushed back on the framing.

The May 2026 announcement is structured differently in three ways. First, the proof was reviewed by named domain experts before publication, not just claimed against benchmark lists. Second, the result is constructive: the model produced an explicit infinite family of arrangements that can, in principle, be independently checked. Third, the construction crosses fields, importing algebraic number theory into a combinatorial geometry question. That cross-field bridge is what serious mathematicians flagged as the genuinely novel content.

What This Means If You Build With Reasoning Models

For engineering teams the most useful takeaway is not that frontier models can now do mathematics, which has been true at varying levels of credibility for two years. The takeaway is that frontier reasoning systems are beginning to produce artifacts that can survive serious expert scrutiny in narrow technical domains. That is the bar the field has been pushing toward, and it is the same bar that matters for agentic systems doing long-running technical work: code refactors, scientific data analysis, hardware synthesis, formal verification.

Two implications worth internalizing:

1. Verifiers are the bottleneck. The reason this result lands while the 2025 results bounced is that there is a clean, mechanically-checkable proof object. If your agent's outputs cannot be cheaply verified, the model's reasoning quality does not matter; you cannot trust the outcome. Invest in the verifier as much as the agent.

2. Cross-domain priors are unlocking. The novel content here is the bridge from algebraic number theory into combinatorial geometry. Modern reasoning models are trained across both literatures and can sometimes surface connections between them. If you have problem statements that have historically required two specialists, give them to a reasoning model with a tool harness and verifier.