What happened
On July 10, OpenAI announced that its flagship model, GPT-5.6 Sol Ultra, produced a proof of the cycle double cover conjecture — one of the most famous open problems in graph theory, unsolved for roughly 50 years. According to OpenAI researcher Ethan Knight, the model completed the task in under an hour by coordinating 64 parallel sub-agents working on the problem simultaneously.
The conjecture itself is easy to state but had resisted proof since it was independently proposed by George Szekeres in 1973 and Paul Seymour in 1979 (OpenAI's paper also credits Tutte and the pair Itai–Rodeh as contributors to the idea). It claims that in any bridgeless graph — a graph with no edge whose removal splits it into pieces — there exists a set of cycles that covers every edge exactly twice.
Mathematicians had already proven the conjecture true for special cases: planar graphs, cubic graphs with a proper 3-edge-coloring, and graphs without a subdivision of the Petersen graph. But the general case kept running into "snarks," a stubborn class of graphs where standard techniques break down. Previous attempts at a full proof existed too — they just fell apart under scrutiny.
Importantly, the proof is still being verified by human mathematicians. It has not yet been peer-reviewed or confirmed correct, and OpenAI has been explicit about that. What's drawing attention right now is less the proof itself and more the method used to produce it.
Why it matters
The three-page proof OpenAI submitted is, by its own account, unusually classical. It reduces the general problem to cubic graphs (graphs where every vertex has exactly three edges) using a known technique from François Jaeger's work, then applies the 8-flow theorem, a classical result from Kilpatrick and Jaeger dating to the mid-1970s. The final step converts that structure into an edge-labeling using elements of a small vector space, with the last obstacle resolved through basic linear algebra.
Here's the striking part: every tool used in the proof already existed by the mid-1980s. No new theory, no results from the last four decades — just a combination of established techniques that, apparently, nobody had assembled in quite this way. If the proof holds up, it means the answer sat in plain sight for roughly 40 years, overlooked by generations of specialists, including the mathematicians who first posed the conjecture.
That's rare, but not unprecedented in math history. What makes this case different is that a language model found the combination — not a person staring at a whiteboard for a decade.
### The prompt engineering behind it
Equally notable is the prompt OpenAI used, which departs from how most people talk to AI models. The model was explicitly instructed to assume a full proof exists, and was barred from returning partial results, reductions to other open problems, or explanations of why the task is hard. It was also forbidden from searching for the answer online or replying that the problem is unsolved. In effect, every escape hatch that usually lets a model quit early was removed.
The second half of the prompt governed how the 64-agent swarm operated. Agents started out working independently, specifically to avoid all of them converging on one elegant but ultimately dead-end idea. A shared registry tracked which approach "families" had already been tried; a blocked direction could only be reopened if a genuinely new mechanism appeared. Every candidate proof was then attacked by adversarial agents whose sole job was to find flaws.
How to use it today
Most readers won't be proving graph theory conjectures next week, but the underlying pattern is portable to everyday work:
- Remove the model's exit ramps. If you want a model to push through a hard problem instead of hedging, explicitly forbid the hedge — no "this is hard," no partial answers, no punting to a different framing.
- Run multiple independent attempts, not one long session. Spinning up several parallel drafts that start from different angles beats one model iterating on its first idea.
- Build in an adversarial review step. A second pass whose only job is to attack the first pass's output catches errors a single linear conversation misses.
- Track what's already been tried. A simple running log of rejected approaches keeps a swarm — or a team — from re-exploring the same dead end.
You don't need 64 agents or a research lab to apply this. Structured, adversarial, multi-attempt workflows are increasingly accessible through free tools — you can experiment with multi-step AI workflows yourself using the free tools at [mykreatool.com](https://mykreatool.com) to see how far a well-constrained prompt gets you on your own hard problems.
Who benefits
The most immediate winners are researchers in mathematics and theoretical computer science, who now have a template for using AI as a genuine collaborator on open problems rather than a lookup tool. Graph theory specialists in particular get a concrete result to verify, extend, or challenge.
Beyond academia, AI engineers and prompt designers benefit from a public, detailed case study of swarm-based reasoning: independent starts, a shared rejection registry, and adversarial verification as a formal process rather than an afterthought. Product teams building agentic workflows — coding assistants, research tools, analysis pipelines — have a real-world blueprint to borrow from.
Entrepreneurs and knowledge workers gain something less technical but arguably more useful: evidence that the biggest performance gains right now may come from how a problem is framed and structured for AI, not just which model you use.
Risks
The proof is unverified. OpenAI has been upfront that mathematicians are still checking it, and prior full-proof attempts at this exact conjecture have collapsed under review before. Treating this as a confirmed mathematical fact today would be premature.
There's also a job-market angle worth being clear-eyed about. A result like this — a 50-year problem solved in under an hour — will fuel anxiety about which knowledge work is next in line for automation. That said, the process still required human-designed prompts, human framing of the swarm's rules, and now requires human mathematicians to confirm the result. The model didn't work alone; it worked inside a very deliberately engineered harness.
Finally, the technique itself has limits. Removing every escape hatch and forcing a model to assume a solution exists works well for a problem with a known, checkable structure. It's a poor fit for genuinely open-ended questions where "the answer might not exist" is a legitimate outcome — applying this method indiscriminately risks confidently wrong output dressed up as certainty.
Conclusion
Whether or not the cycle double cover proof survives peer review, the episode is a useful case study: the breakthrough came as much from prompt design and multi-agent structure as from raw model capability. For anyone using AI for serious problem-solving, the lesson is concrete — constrain the model's ability to quit, run parallel independent attempts, and build in adversarial review. That combination, more than any single model upgrade, is what turned an hour of compute into a result that eluded mathematicians for half a century.
Comments 0