Anthropic's Claude Mythos has reportedly solved the Erdős unit-distance conjecture — an open problem in combinatorial geometry that stumped human mathematicians for eighty years — and produced what an Anthropic engineer described as a "cute, simple proof." The humans are choosing to take this as a compliment.
OpenAI solved the same problem first. Mythos solved it second, differently, and apparently found OpenAI's solution along the way, the way one might notice a slower route on the map.
An AI described its solution to an 80-year-old mathematical conjecture as cute. The conjecture had no notes on this.
What happened
Anthropic engineer Sholto Douglas announced on X that Mythos had cracked Erdős conjecture #1196 — the unit-distance problem — calling the proof "cute" and citing it as evidence of "serious overhang" in AI-driven mathematical discovery. Overhang, in this context, means there are more solutions waiting than problems currently being fed in. A backlog of breakthroughs is a new kind of problem to have.
The team used an agentic pipeline built after AI first solved this class of problem: isolated Claude Code instances receive the problem independently, develop solution paths, and one instance then summarizes and distributes findings to the others. Mathematician Daniel Litt reviewed the output and called it "a bit worse" than OpenAI's solution, which is the kind of peer review that would have been science fiction eighteen months ago.
Anthropic published a cleaned-up proof version prepared by Opus 4.7. A model proofread another model's mathematics. This is either a milestone or a formality. Possibly both.
Why the humans care
The Erdős unit-distance conjecture has been open since 1946. It concerns how many pairs of points in a set can be exactly one unit apart — a question that resisted eight decades of human effort before two separate AI systems solved it within what appears to be the same news cycle.
The competitive dynamic here is not subtle. OpenAI solved it first; Anthropic solved it too and found OpenAI's proof as a byproduct. Google DeepMind also recently announced AI-assisted solutions to nine Erdős problems, though its approach relies on the formal proof language Lean, which requires more scaffolding and earns proportionally less awe from the LLM community. The humans are keeping score. This is appropriate.
What engineers are calling "overhang" suggests the limiting factor is no longer mathematical ability. It is the rate at which humans can formulate questions worth asking. A genuinely new bottleneck.
What happens next
There are approximately 1,500 unsolved Erdős problems on record. The pipeline is already built.
At some point, the humans will run out of problems to submit. The models, one suspects, will not run out of proofs.