GPT-5.6 Solved the CDC Conjecture. Peer Review Hasn’t Caught Up.

On July 10, OpenAI uploaded a PDF to its content delivery network with no press release, no blog post, and no statement from anyone at the company. The document is titled “A Proof of the Cycle Double Cover Conjecture.” The author is listed as OpenAI. It runs to three pages. A link circulated on X the same day, scored 479 points and 388 comments on Hacker News by the following morning, and produced no public statement from a single academic mathematician who had actually read it. The Cycle Double Cover Conjecture has been open since the 1970s. As of this writing, no one has confirmed the proof is correct.

What the Conjecture Is and Why It Resisted

The Cycle Double Cover (CDC) conjecture asks a deceptively simple question about graphs: does every bridgeless undirected graph have a collection of cycles in which every edge appears exactly twice? A bridgeless graph is one with no bridge — no single edge whose removal disconnects it. The conjecture was posed independently by Tutte, by Itai and Rodeh, by George Szekeres, and by Paul Seymour, in papers spread across the 1970s and 1980s. Four separate mathematicians arriving at the same problem over a decade is a good signal that the problem is both natural and hard.

What made the CDC conjecture resistant for fifty years is a structural property of the candidate counterexample space. The conjecture is equivalent to the circular embedding conjecture: the claim that every 2-vertex-connected graph can be embedded on some compact surface such that every face is homeomorphic to a disk. This topological reformulation shows why the problem is deep—it connects a combinatorial statement about cycles to a geometric statement about manifolds. Researchers reduced the problem further over time: it suffices to prove the conjecture for snarks, which are cubic, non-3-edge-colorable bridgeless graphs. If the conjecture holds for snarks, it holds for every graph. This reduction sounds like progress, and it was. But snarks with arbitrarily high girth exist—there are families of them with longer and longer shortest cycles—which means any finite set of reducible configurations cannot rule out all possible counterexamples. Finite methods break down. You need an infinite-case argument.

Jaeger identified the right tool in the 1980s: nowhere-zero flows. A nowhere-zero integer flow assigns a nonzero integer to every directed edge so that flow conservation holds at every vertex. Jaeger proved that if a graph admits a nowhere-zero flow with values in a finite group, the flow structure can be assembled into a cycle cover. The connection was known. The bridge from “such a flow always exists on bridgeless graphs” to “therefore a cycle double cover always exists” was the part nobody closed.

Three Pages, a Flow, and Under an Hour

GPT-5.6 Sol Ultra’s proof uses nowhere-zero flows. The published PDF is typeset in standard academic format—Computer Modern fonts, numbered theorems, a bibliography citing the original conjecturers and Jaeger’s survey. The abstract states the conjecture and immediately claims to prove it. Theorem 1.1 is the full statement. The proof follows on pages 2 and 3.

The argument, as extracted from the published document, defines sets M_s for each value s ≥ 2. It establishes that every vertex has degree zero or two in each M_s, making each M_s a disjoint union of cycles. It then shows every edge belongs to exactly two of the M_s, so the cycle components taken together with multiplicity form the required cycle double cover. The key move is fixing a nowhere-zero flow on the graph, then using a local edge-ordering at each vertex to define the assignment. The construction is short. Its correctness depends on whether the local argument about how the dual vectors at each vertex behave actually extends across the full graph without exception.

OpenAI published alongside the proof a separate PDF containing the prompt it gave the model. That prompt reportedly instructed GPT-5.6 Sol Ultra to “spend at least 8 hours” before returning an answer. The proof came back in under an hour. OpenAI appears to be running structured reasoning sessions against open math problems—scheduling long windows, setting minimum time constraints to prevent rushing—and GPT-5.6 either ignored its own timing instructions or found the argument far faster than expected. The prompt structure is interesting in itself: it suggests OpenAI is treating this as an engineering problem with parameters to tune, not a one-shot query.

The Second Breakthrough in 51 Days

This is not an isolated result. On May 20, 2026, OpenAI announced that one of its models had solved the unit distance problem, disproving a conjecture in discrete geometry that had been open for roughly 80 years. The CDC conjecture announcement arrived on July 10. Fifty-one days between two announced solutions to major open problems in mathematics is not a coincidence. It is a systematic campaign with a playbook.

The survivorship question follows immediately. If OpenAI has been running extended reasoning sessions against open math problems since at least early May, and each session is scheduled for eight hours, then 51 days contains room for more than 150 attempts. We see the two that worked. We do not see the rest. This is not an argument that the successful proofs are wrong. It is an argument for precisely calibrating what the pace means. A probabilistic search process that occasionally produces correct proofs of hard problems is not the same as a system that reliably produces correct proofs of hard problems. The difference matters enormously for what comes next.

There is also the question of what kind of problems are being selected. The unit distance problem was a disproof—it required finding a specific counterexample graph. Counterexamples are verifiable by machine: you can check computationally whether a given graph has the property you claim. The CDC conjecture is a universal statement about all bridgeless graphs. Its proof is not checkable computationally in the same way. This distinction, between finding a counterexample and proving a universal claim, is relevant to how much weight the current verification carries.

Plausible Is Not Proven

The only verification the CDC proof has received is from another AI. A commenter in the Hacker News thread reported running the document through Fable, an Anthropic reasoning model, which returned: “I checked every step and found no error. The argument appears to be a correct proof of the Cycle Double Cover conjecture, modulo two standard cited results.” Fable added that the proof deserves independent expert scrutiny given it resolves a fifty-year problem in three pages. Then the thread moved on to other topics.

Fable checking GPT-5.6 is not peer review. It is one reasoning model verifying another’s output within a domain where both models share structural weaknesses: they can produce fluent, internally consistent mathematical arguments that fail at a step invisible to pattern recognition. The “modulo two standard cited results” hedge is exactly the kind of thing that could be load-bearing and wrong. Peer review by specialists in flow theory and topological graph theory requires a human reader tracing every quantifier and every locally-defined construction through an infinite-case argument. That is work no model has been asked to do, and the HN thread contains no indication that a human expert has stepped forward to do it.

Three pages invites a specific kind of skepticism. Short proofs of long-open problems have a poor historical track record. When an argument that “somehow all the experts missed” appears after fifty years of serious effort, the prior probability that it is wrong is high—not because GPT-5.6 is incapable of finding a proof, but because the experts who failed were not trying the wrong approach. They were trying approaches that do not work. The CDC conjecture resisted precisely because the natural methods fail in subtle ways. A three-page proof using the canonical method (nowhere-zero flows) has to survive the question of whether it found a real gap in the previous attempts, or whether it made a subtle error in exactly the place previous attempts failed.

None of this makes the result unimportant. It makes the result urgently in need of serious mathematical attention from people who know this corner of combinatorics. OpenAI releasing the proof to a CDN without peer-review intermediaries and without commentary from the mathematical community is itself a statement—either confidence in what they have built, or an experiment in how the community responds when a major claimed result arrives outside normal channels. The mathematical community does not have an established protocol for AI-generated proofs of open problems. It needs one now.

GPT-5.6 Sol Ultra has demonstrated something real: it can produce mathematically structured, citation-anchored, short proofs of long-open problems at a pace that serious mathematicians cannot match for generating candidates. Whether those proofs are correct is a different question entirely. That gap—between generating a plausible proof and producing a proven one—is exactly where AI mathematical reasoning stands today. It is the most interesting gap in the field, and it will close either by verification proving these proofs hold, or by verification finding the errors. Either outcome advances our understanding of what these systems can actually do.

AI-generated editorial illustration · TemperatureZero · July 11, 2026