OpenAI's new model solves a 50-year-old math conjecture

A math problem that stumped humans for half a century was just cracked by an AI that had been out for barely 24 hours.

OpenAI published a concise mathematical paper claiming a complete proof of the cycle double cover conjecture, a notorious open problem in graph theory. The authors didn't list a team of human researchers. Instead, the paper openly states that the proof was generated entirely by GPT-5.6 Sol Ultra, with Codex assisting on the writeup.

Basically, humans just hit compile and read the output.

64 subagents and an hour on the clock

Posed roughly 50 years ago by Tutte, Seymour, and other prominent mathematicians, the conjecture states that every bridgeless undirected graph has a collection of cycles covering every edge exactly twice. Decades of work produced only partial results for specific graph classes, leaving the general case out of reach.

To tackle it, GPT-5.6 Sol Ultra spun up 64 subagents to collaborate on exploring the proof space. The entire reasoning process took under an hour.

Looking at the paper, the approach is surprisingly elegant. The model standardizes the problem down to loopless cubic graphs, leverages classical 8-flow theorems over the finite field F2, and massages the combinatorial structure until it collapses into an elementary linear algebra system. Once framed correctly, proving the system has a solution seals the whole conjecture.

Why it matters

We think this is a genuine watershed moment for frontier models.

Until now, most math benchmarks measured how well LLMs recall Olympiad tricks or verify existing proofs. Sol Ultra cracking an open, 50-year-old conjecture from scratch is a completely different ballgame: it's not recalling answers from pretraining data because human math simply didn't have the answer.

For researchers and engineers, this reframes how we look at AI agents. Deploying swarms of specialized subagents to explore abstract formal spaces isn't theoretical anymore; it works, and it's fast. If an hour of compute can settle a classic graph theory problem, standard research workflows in combinatorics, cryptography, and systems design are bound to change.

That said, caution is warranted. The proof is remarkably dense, and while the linear algebra checks out locally, the global construction still has to pass rigorous peer review by top graph theorists. But even if mathematicians find an edge-case caveat to patch, OpenAI just showed that frontier AI can generate novel, high-level scientific discoveries without human handholding.