GPT-5.6 Sol claims a 50-year math proof
OpenAI claims GPT-5.6 Sol Ultra proved the Cycle Double Cover Conjecture using 64 subagents in under an hour. Mathematicians are reviewing; no peer review yet.
OpenAI published a PDF on July 10 claiming GPT-5.6 Sol Ultra produced a complete proof of the Cycle Double Cover Conjecture, a graph theory problem open since George Szekeres posed it in 1973 and Paul Seymour reformulated it in 1979. The conjecture states that every bridgeless graph has a collection of cycles in which each edge appears exactly twice.
The system used 64 parallel subagents coordinated by a planning model. Total time: under one hour. OpenAI published both the proof and the prompt, one day after GPT-5.6 Sol Ultra became generally available.
Mathematicians are reading the proof publicly. No formal peer review has happened. The Cycle Double Cover Conjecture has attracted several claimed proofs over the decades that turned out to have gaps, so expert informal verification is the right first step before treating this as settled.
What it means for you
The architecture is what matters here: 64 agents working in parallel on subproblems, coordinated at the planning level, not one model grinding through the problem alone.
If a 64-agent swarm can hold enough mathematical structure to produce a claimed proof of a 50-year open conjecture, the same design applies to any domain where large problems can be broken into parallel tracks. Regulatory analysis across jurisdictions, drug interaction modeling, distributed debugging on large codebases all fit the same shape. The math conjecture is the proof of concept. The coordination pattern is what generalizes.
One caveat that matters: OpenAI self-reported this and has obvious incentives to announce it. “Claimed proof” is where this sits until mathematicians verify it independently. Until then, treat the result as promising and the architecture as worth studying. For a deeper look at what this pattern means for agent design, see what an AI math proof means for your agents.