OpenAI Releases Prompt Behind GPT-5.6 Sol Ultra Math Proof

The interesting artifact here is not the proof, it is the prompt. OpenAI put a short PDF on its CDN titled "Prompt Used for A Proof of the Cycle Double Cover Conjecture", a rare look at the actual instructions handed to a frontier model to get it to produce a claimed solution to a fifty-year-old open problem in graph theory.

The setup, as The Decoder reported, is GPT-5.6 Sol Ultra with 64 subagents running in parallel, arriving at a three-page proof in under an hour. The Cycle Double Cover Conjecture was formulated independently by several mathematicians in the 1970s, and the prompt itself does something worth staring at. It tells the model to assume a complete proof exists, bans it from searching the internet, bans it from answering that the conjecture is unsolved, and instructs it to compute for at least eight hours before it can even consider giving up. It also spins up adversarial agents whose job is to find holes in candidate proofs.

Why this matters more than a typical model announcement is that a lab has published, alongside a headline capability claim, the exact recipe that produced it. If the technique holds up it becomes a template that other groups will copy, long-horizon reasoning as a scaffolded ensemble with a hostile red-team baked into the same prompt.

The honest caveat is that the proof itself is not yet accepted. There is no Lean or Coq formalization, no peer review, and mathematician Thomas Bloom has already flagged that OpenAI's paper does not cite prior work in the area, a pattern he called a frequent issue with AI-generated papers. The reporting also does not tell you how much human editing sat between the raw multiagent output and the tidy three-page write-up, or whether the adversarial-agent setup generalizes to other open problems or was tuned to this one.

The direction to watch is not whether this particular proof survives contact with the community. It is whether "assume the answer exists, run adversaries against candidate answers, do not stop for hours" becomes the standard recipe for pointing frontier models at hard problems. If it does, the prompt will end up mattering more than the proof.