Eighty years is a long time for a geometry problem to stay open. Erdős posed the planar unit distance conjecture in 1946, and it sat there, deceptively simple to state, stubbornly resistant to resolution, until OpenAI announced last week that one of its reasoning models had disproved it.
We covered the announcement on May 21. What’s changed since: the mathematical community is responding.
Noga Alon of Princeton, quoted directly in OpenAI’s announcement, called it “one of Erdős’ favorite problems.” That’s not a dismissal framed as praise, Alon is a combinatorialist of the first rank, and his characterization matters. Mathematician Gil Kalai, whose blog commentary has circulated widely since the announcement, addresses the result directly: the disproof is real and the mathematics checks out. External academic engagement at this level is notable. This wasn’t a proof handed to a credulous press office.
The result itself: per the proof paper (arXiv:2605.20695, once accessible), the model established that the maximum number of unit-distance pairs in a set of n points grows faster than Erdős conjectured, specifically, that for a fixed δ > 0, ν(n) ≥ n^(1+δ) infinitely often. OpenAI describes the problem as “one of the best-known questions in combinatorial geometry, easy to state and remarkably difficult to resolve,” citing the 2005 reference text *Research Problems in Discrete Geometry*. That’s accurate. The problem’s difficulty came not from obscurity but from the gap between how natural the question is and how hard the answer proved to be.
The model is described as a general-purpose reasoning model using extended reinforcement-learning-driven chain-of-thought. According to OpenAI, the proof was developed in collaboration with external mathematicians, reportedly including Thomas Bloom and Timothy Gowers. Those co-author names come from OpenAI’s announcement, the arXiv paper, once accessible, is the authoritative record. The proof reportedly runs to approximately 125 pages of chain-of-thought output.
The part nobody mentions in the press coverage: OpenAI hasn’t disclosed which model this is. The informal label circulating in developer communities (“GPT-next”) doesn’t appear anywhere in OpenAI’s announcement. What’s confirmed is that the model is general-purpose, not a dedicated theorem prover, not a system built specifically for discrete geometry. That distinction matters more than the name.
What to watch
the arXiv paper (2605.20695) becoming fully accessible is the next verification milestone. That’s where the mathematical record lives, the co-authorship, the full proof structure, and the formal notation. Peer review of a proof at this scale takes time. The Kalai blog commentary is a positive early signal, not a final verdict. OpenAI has also noted that Erdős offered a monetary prize for resolving this problem; the amount wasn’t confirmed in available materials.
Analysis
The model identity has not been confirmed by OpenAI. Community labels like 'GPT-next' don't appear in the official announcement. What's confirmed is the result and the methodology, general-purpose RL reasoning, not a specialized solver. That's the claim with the longest tail.
The practical question for anyone building with AI in research contexts: this result was produced by a general-purpose model, not a specialized math system. That changes the prior. If RL-driven chain-of-thought can generate 125 pages of valid combinatorial reasoning in an open problem that stymied human mathematicians for eight decades, the relevant variable isn’t “can AI do hard math” anymore. The question is scope, what categories of mathematical work are now within reach, and what does that mean for how research teams should think about AI assistance?
TJS synthesis
Don’t treat this as a stunt. The external mathematician engagement, Alon quoted, Kalai engaging in substantive commentary, places this result in a different category from vendor-produced benchmark claims. The arXiv paper is the document to read when it resolves. Until then, the verified story is: a general-purpose reasoning model produced a result in combinatorial geometry that external mathematicians are taking seriously. That’s the signal worth tracking.