GPT-5.6 Sol Proves a Theorem Humans Couldn't Crack for 50 Years

OpenAI's GPT-5.6 Sol Ultra built a complete proof of the Cycle Double Cover Conjecture — an open graph theory problem dating to the 1970s — in under an hour. Sixty-four subagents ran in parallel, and the model was deliberately kept unaware that the problem was considered unsolved.

Mathematician Thomas Bloom called the proof "short and elementary — the kind that could have appeared back in the 1980s." One critique: the proof does not cite the foundational 1983 paper by Bermond, Jackson, and Jaeger, though it clearly builds on their ideas. This is a recurring issue with AI-generated proofs.

The key question is no longer "can AI do serious math?" but "what exactly is it doing?": finding genuinely new paths, or cycling through known tools with unprecedented persistence. The answer changes how we should think about AI in science — not as a replacement for the scientist, but as a machine that pushes problems to completion where humans give up.