OpenAI’s GPT-5.6 Sol Ultra has produced a three-page proof of the Cycle Double Cover Conjecture — a graph theory problem that has remained open for over 50 years, posed independently by Tutte, Itai and Rodeh, Szekeres, and Seymour in the early 1970s. The proof appeared on OpenAI’s CDN on July 10 and sparked intense discussion on Hacker News, where it reached 437 points and 354 comments in half a day.
🔍 THE BOTTOM LINE
If the proof is correct, it’s a landmark moment for AI-assisted mathematics: an open problem that survived five decades of human effort, cracked by a model in a single session. But “if” is doing a lot of work. The proof has not been peer-reviewed. It is three pages long — unusually short for a problem of this caliber. And it arrived the same week OpenAI’s safety leadership collapsed, which means the model that produced it is operating in an environment where the guardrails are visibly loosening. Mathematicians are skeptical. They should be.
What the Cycle Double Cover Conjecture Actually Is
The conjecture is deceptively simple to state: every bridgeless undirected graph has a collection of cycles that covers every edge exactly twice. A “bridgeless” graph is one where no single edge removal disconnects the graph. A “cycle” is a path that starts and ends at the same vertex. The conjecture says you can always find a set of cycles such that every edge in the graph appears in exactly two of those cycles.
That’s it. The statement fits in a sentence. The proof has eluded graph theorists since 1973.
Partial results existed: Jaeger proved it for planar graphs, Szekeres for 3-edge-colorable cubic graphs, and Alspach, Goddyn, and Zhang proved it for graphs with no Petersen subdivision. But the general case — every bridgeless graph, no exceptions — remained open. It is one of the most famous unsolved problems in graph theory, alongside the Erdős–Faber–Lovász conjecture and the graceful tree conjecture.
What GPT-5.6 Sol Ultra Actually Did
The proof document is titled “A Proof of the Cycle Double Cover Conjecture” and carries only the OpenAI logo as institutional affiliation. The statement of AI use is explicit: “The proof in this note is entirely due to GPT 5.6 Sol Ultra and the writeup with Codex (with GPT 5.6 Sol).”
The approach is elegant. The proof reduces to loopless cubic graphs (standard), then uses the 8-flow theorem to label edges with elements of F32 such that the sum at each vertex is zero. The key step is converting this flow labeling into a labeling by two-element sets where each element appears zero or two times at each vertex. This reduces to a linear algebra argument over F2.
The proof’s structure follows a path that human mathematicians had explored: reduce to cubic graphs, use existing flow theorems, find a clever combinatorial construction. The novelty is in the specific construction — the local-to-global argument via Lemma 2.2, which shows a certain system of equations over F32 always has a solution. It’s the kind of move a mathematician would recognize as plausible but non-obvious.
Why Mathematicians Are Skeptical
Three reasons, in order of severity:
1. It has not been verified. No graph theorist has publicly confirmed the proof. The Hacker News discussion includes several commenters with apparent mathematical background flagging potential gaps, particularly around Lemma 2.2 and whether the linear algebra argument fully closes. Three pages is short for a problem that has resisted proof for 50 years — but short proofs are not impossible (the proof of the four-color theorem’s key lemma was relatively compact, even though the full proof was computational).
2. AI-generated proofs have a verification problem. Large language models are known to produce text that reads as mathematically correct but contains subtle errors — hallucinated references, incorrect logical steps buried in dense notation, or gaps disguised by plausible-sounding arguments. The notation in this proof is clean and uses standard results (Tutte’s group-flow theorem, the 8-flow theorem), which makes it harder to spot errors but also harder to dismiss. The proof needs a graph theorist with expertise in flow theory to check it line by line.
3. The timing is suspicious. The proof dropped the same week that OpenAI’s safety head Johannes Heidecke departed, chief futurist Joshua Achiam left, and the company’s own deployment safety page flagged “concerning forms of misaligned behavior” in GPT-5.6. Whether that’s correlation or causation, the proof arrives in a moment when OpenAI’s internal oversight is visibly fraying. The proof itself might be correct — but the institutional context makes trust harder.
What Happens Next
The proof will be scrutinized by graph theorists. If it holds, it will likely appear on arXiv with human co-authors who verified it, and eventually in a journal like the Journal of Combinatorial Theory. If it has a gap, the gap will be found — the graph theory community is small, focused, and has been working on this problem for decades. They will check it fast.
If the proof is correct, the implications go beyond graph theory. It demonstrates that the most advanced reasoning models can tackle open problems in pure mathematics — not just competition problems or exercises, but genuine unsolved conjectures. That’s a different category of capability than writing code or answering questions. It’s the kind of thing that changes how mathematicians work.
If it’s wrong, it’s still a data point: GPT-5.6 Sol Ultra got close enough to a 50-year problem that it took experts hours to find the gap, not minutes. That’s not nothing.
This connects to the broader AI capability arc we’ve been tracking: GPT-5.6’s global launch, Cognition’s SWE-1.7 nearly matching frontier coding models, and the acceleration of model release cycles. Each model generation is reaching further into domains that require genuine reasoning. The CDC conjecture proof — right or wrong — is a signal of where this is heading.
NZ Angle
New Zealand has a small but active mathematics community, particularly in combinatorics and graph theory at universities like Auckland and Victoria. If AI models start producing research-grade proofs, the academic publishing landscape shifts — not just for NZ, but globally. The question for Kiwi mathematicians is the same as for everyone else: do AI tools become collaborators, or competitors? The answer likely depends on whether proofs like this one hold up.
❓ FAQ
What is the Cycle Double Cover Conjecture? Every bridgeless undirected graph has a collection of cycles that covers every edge exactly twice. It was posed in the early 1970s by multiple mathematicians independently and has been open for over 50 years.
Has the proof been verified? No. The proof appeared on OpenAI’s CDN on July 10, 2026. It has not been peer-reviewed or confirmed by independent graph theorists. The Hacker News discussion includes skepticism about Lemma 2.2.
What is GPT-5.6 Sol Ultra? It is OpenAI’s most powerful reasoning model variant, capable of extended chain-of-thought reasoning. The proof document states it was produced entirely by GPT-5.6 Sol Ultra, with the writeup formatted using Codex.
Why does the timing matter? The proof was released the same week OpenAI’s safety head and chief futurist both departed, and the company flagged “concerning misaligned behavior” in GPT-5.6. The institutional context doesn’t affect the proof’s correctness but does affect trust.
Could a three-page proof really solve a 50-year problem? Yes, in principle. Short proofs of long-standing problems exist — the proof of the Pólya conjecture was eventually disproved, but genuine short proofs do appear. The brevity warrants scrutiny, not dismissal.
🔍 THE BOTTOM LINE
GPT-5.6 Sol Ultra produced a proof that, if correct, solves one of the most famous open problems in graph theory. The proof is elegant, uses standard tools in a novel way, and is short enough to check in an afternoon. It is also unverified, produced by a model whose safety oversight is collapsing in real time, and dropped into a community that has been burned by AI-generated proofs before. The right response is neither celebration nor dismissal. It’s to read the proof, check Lemma 2.2, and wait for a graph theorist to render judgment. The math will either hold or it won’t. Everything else is noise.
