Claude Mythos, Anthropic’s frontier reasoning model that we have been tracking through its closed-consortium phase and the imminent Mythos 1 general release, reportedly produced a working disproof of the same planar unit-distance Erdos conjecture that OpenAI’s reasoning model resolved earlier this month, The Decoder reported on May 26.
Mythos found a proof reviewers described as “cute and simple,” meaning shorter and more elegant than OpenAI’s original. Separately, Mythos was also reportedly able to find OpenAI’s solution as an alternative approach, demonstrating that the model is not stuck on a single proof strategy. Anthropic’s piece notes that Mythos’ shorter proof was judged slightly weaker than OpenAI’s on technical depth but is the kind of result that would have stood alone if Mythos had been the first to publish.
The reading worth taking from this: when frontier reasoning models converge on the same open mathematical problem from independent training runs and produce qualitatively different working proofs, the field has moved past “AI can solve math” as the headline finding. The question now is which model produces the cleanest argument, and that question maps onto the same Solver-versus-Creator gap BenchBench identified earlier this week. Different models find different proofs, and the diversity of approach is a feature rather than a redundancy.
Reported by The Decoder on 2026-05-26.
