Claude Mythos has an Erdős proof claim; verification is the real test

The Decoder reports that Anthropic’s Claude Mythos has reportedly solved Erdős’ unit-distance problem, shortly after OpenAI announced its own breakthrough on the same mathematical terrain. According to the source summary, engineer Sholto Douglas says Mythos cracked the 1946 conjecture “over the weekend” with a “cute, simple proof”.

The important point is not only that this is a known problem in discrete geometry, but that the same target appears to be emerging as a test case for multiple frontier AI systems at almost the same time. The Erdős unit-distance problem asks how many pairs of points at distance one can occur among n points in the plane. Problems like that usually do not fall to blind brute force; they require structure, proof strategy, and a feel for edge cases.

That is why the phrase “cute, simple proof” matters more than the victory label. In mathematics, a simple proof often means someone found the right angle of attack, not merely a longer computational derivation. If Anthropic’s model really produced such a proof, the question is no longer only whether AI can follow existing literature, but whether it can generate explanations that humans can inspect, shorten, and fold into normal mathematical work.

There is still a necessary caution here. The available account says “reportedly”; it does not provide the full proof, peer review, an official paper, or formal verification in the supplied context. That means the claim should be read as an early signal, not as a closed scientific fact. In mathematics, a model does not earn trust by saying it has a proof. The proof has to survive expert reading, formalization, or at least careful human checking.

The larger implication sits in Douglas’ description of “serious overhang” in AI-driven math discovery. In plain terms, model capability may already be ahead of the ways it is routinely measured. If these systems are rarely aimed at concrete open problems, some of that capacity remains invisible until someone activates it with the right prompt, environment, or aggressive experiment.

OpenAI’s role in the story sharpens the industry angle. If OpenAI and Anthropic are approaching the same mathematical breakthroughs, academic mathematics is getting a new kind of competitor and collaborator: models that can rapidly search the space of ideas, while still needing hard verification. That is not the end of mathematicians. It is a shift in where the first draft of a proof may appear, and who has to take it seriously first.