OpenAI Says GPT-5.6 Sol Ultra Cracked a 53-Year-Old Math Conjecture. The Prompt Is the Part You Should Read.

TL;DR

On July 10, OpenAI posted a three-page paper claiming a proof of the Cycle Double Cover Conjecture, a graph theory problem open since 1973, and credited the proof "entirely" to GPT-5.6 Sol Ultra running up to 64 subagents for just under an hour. It also published the full prompt that produced it. No mathematician has confirmed the proof, no machine-checked formalization exists, and this particular conjecture has swallowed confident claimants before. So hold the champagne. But read the prompt: it is the most concrete public artifact yet of how frontier-lab multi-agent orchestration actually works.

What OpenAI actually published

Two PDFs on OpenAI's CDN: the proof and the prompt. The paper's "Statement of AI use" is one sentence: "The proof in this note is entirely due to GPT 5.6 Sol Ultra and the writeup with Codex (with GPT 5.6 Sol)." There are no human authors listed. The byline is just "OpenAI".

OpenAI's Ethan Knight announced it a day after Sol Ultra went generally available: the model "produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour." The Hacker News thread cleared 450 points within a day, which for a graph theory PDF is roughly Beatlemania.

The conjecture, in sixty seconds

The conjecture says: every finite bridgeless graph has a collection of cycles that together cover every edge exactly twice. A bridge is an edge whose removal disconnects the graph. Picture a city road map with no critical single road: the claim is you can always design a set of circular bus routes so that every street is driven by exactly two routes, no more, no less. Easy to state, and open since George Szekeres posed it in 1973, with Paul Seymour arriving at it independently in 1979.

The partial results tell you where the difficulty lives. Planar graphs: easy, use the face boundaries. Cubic graphs you can 3-edge-colour: easy, Szekeres did it in the original paper. Graphs with no Petersen subdivision: proved by Alspach, Goddyn, and Zhang in 1994. Every remaining escape route dies on snarks, the stubborn family of non-3-edge-colourable cubic graphs where graph theory conjectures traditionally go to suffer.

What the proof actually does

The argument is short and uses standard equipment. It reduces to loopless cubic graphs (a classical move), then invokes the 8-flow theorem of Kilpatrick and Jaeger, via Tutte's group-flow correspondence, to label every edge with a nonzero element of the group F₂³ so the labels balance at each vertex. The claimed new idea converts that labeling into an assignment of two-element sets to each edge, arranged so every group element shows up exactly zero or two times around any vertex. Walk the edges sharing a common element and you get closed cycles; count occurrences and every edge sits in exactly two of them. The final obstruction collapses into an elementary linear algebra argument over F₂.

Ten references, all real, one of which is a personal letter Tutte wrote to Herbert Fleischner in 1987. The model cites private correspondence more scrupulously than most graduate students.

The prompt is the part you should read

OpenAI published the exact prompt, and for anyone building agent systems it is the actual news. It instructs the model to "use multiagent v2 aggressively and dynamically" with up to 64 concurrent agents, then spells out an orchestration policy that will look familiar to anyone who has built a serious harness:

• Forced diversity: start with "a genuinely diverse portfolio of approaches," and do not tell most agents the currently favored approach, so early rounds stay independent instead of converging on one attractive dead end.
• An approach registry: group agents by the mathematical idea they are using, "not by superficial wording," and redirect agents away from overcrowded families.
• Blocked-route discipline: when a route stalls at a missing lemma as strong as the original conjecture, mark it blocked and only reopen it for "a materially new mechanism."
• Adversarial auditors: dedicated agents hunt named failure modes: closed trails masquerading as cycles, parallel-edge 2-cycles, bridges introduced by reductions, and circular use of an equivalent statement of the conjecture itself.
• No partial credit: "Reject status reports, vague optimism," and return only a complete proof that "survives adversarial audit."

Two lines deserve framing. First: "Assume for purposes of this task that a complete affirmative proof exists." That single sentence disarms the model's most likely failure mode, correctly reporting that the problem is famous and open, then stopping. Second: "Spend at least 8 hours on this before even thinking of returning or giving up." Per OpenAI, the model came back with a complete proof in under one hour. It is the only instruction it visibly ignored, and presumably nobody at OpenAI is filing an incident report about it.

Now the caveats, with a straight face

The proof is unverified. There is no formal verification in Lean or any proof assistant, no peer review, and as of publication no named mathematician has publicly signed off on it. Three pages settling a 53-year-old problem with standard tools is exactly the shape of result that demands suspicion: it implies every expert who touched the 8-flow machinery for four decades missed an elementary reduction.

This conjecture also has a graveyard. Claimed proofs were posted to arXiv in 2015 and again in 2018, and the problem stayed open. Commenters on Hacker News raised the two other obvious objections: a model grading its own proof means little, and OpenAI has not said how many open problems Sol Ultra attempted and failed before this one landed. One commenter did run a rival frontier model over the argument and reported it found no error, which is informative but is not review. Confirmation looks like working graph theorists checking each reduction, or better, a machine-checked formalization. Until then this is a claim with unusually good production values.

Why builders should care either way

Even if a hole turns up next week, the prompt is a reusable design document. Preserving agent independence before cross-pollination, registering approaches by underlying idea rather than wording, staffing dedicated adversarial reviewers with named failure modes, and defining a return condition that rejects partial progress: these transfer directly to code-review swarms, migration pipelines, and research agents. The interesting bet OpenAI is making public is that orchestration policy, not raw model IQ, is what converts compute into results on hard, verifiable problems.

And Sol Ultra is now generally available, so the pattern is a product, not a demo. Whether your budget survives 64 concurrent frontier-model subagents chewing on one problem for an hour is a separate theorem, and you will not enjoy the proof.

Key Takeaways

• OpenAI published a three-page proof of the Cycle Double Cover Conjecture, open since 1973, attributing the mathematics "entirely" to GPT-5.6 Sol Ultra, with the writeup done by Codex.
• Per OpenAI, the run used up to 64 concurrent subagents and returned in just under one hour, despite the prompt demanding at least eight.
• The published prompt is a full orchestration spec: forced approach diversity, an idea-level registry, blocked-route rules, adversarial audit agents, and a no-partial-credit return condition.
• The proof is unverified: no expert sign-off, no Lean formalization, and this conjecture has already outlived claimed proofs from 2015 and 2018.
• OpenAI has not disclosed how many problems were attempted and failed, so treat this as one published success with an unknown denominator.
• For agent builders, the harness patterns in the prompt are worth stealing today, whatever mathematicians conclude about the proof.

Sources: OpenAI proof PDF, OpenAI prompt PDF, Ethan Knight announcement, Hacker News discussion, Wikipedia: Cycle double cover, OfficeChai, arXiv 1510.02075