← Back to feed

The Cycle Double Cover Conjecture (CDCC) has been one of the most important open problems in graph theory for over 50 years. This morning, OpenAI’s GPT-5.6 Sol Ultra appears to have solved it.

The proof PDF, published directly at cdn.openai.com, represents OpenAI’s third publicly documented AI-generated mathematical breakthrough in 2026: GPT-5.2’s gluon amplitude result in February, an internal model’s disproof of the planar unit distance conjecture in May, and now the CDCC via the commercial GPT-5.6 Sol Ultra.

What the Conjecture States

The CDCC, posed independently by Szekeres in 1973 and Seymour in 1979, states that every bridgeless graph has a cycle double cover: a collection of cycles such that each edge appears in exactly two of them. A graph is bridgeless if removing any single edge leaves it connected.

The statement is easy to picture. Polyhedral graphs — the graphs of convex polyhedra — trivially have cycle double covers via their face boundaries: each edge borders exactly two faces. The conjecture asks whether this holds for every bridgeless graph, however irregular or pathological.

Why It Matters

The CDCC is not an isolated problem. It sits at the intersection of several major unsolved questions in combinatorics and topology.

It implies the Circular Embedding Conjecture: that every 2-edge-connected graph can be circularly embedded on some surface, with every face a cycle. The stronger orientable version of the CDCC, if the proof addresses it, would imply Tutte’s nowhere-zero 5-flow conjecture — itself one of the most studied open problems in graph theory.

Previous proof attempts have all failed. Papers claiming proofs (arXiv 1510.02075 by Radcliffe, 1811.08719, 1202.0569 and others) were each found to contain errors upon close reading, with none surviving the mathematical community’s scrutiny. Through computer search, it is known that any minimal counterexample must have girth at least 12 — ruling out thousands of specific graph families — but the general case remained open. Reducing the conjecture to cubic (3-regular) graphs is possible via standard arguments; the hard part is proving it for those.

The Third AI Math Breakthrough of 2026

The pattern across OpenAI’s three 2026 breakthroughs is structurally different each time. GPT-5.2 in February worked alongside human mathematicians at IAS and Harvard on the gluon amplitude problem, proposing a formula that humans then proved. The unit distance disproof in May came from an internal scaffolded model evaluated against a collection of Erdős problems, with the proof checked by external mathematicians before announcement. This CDCC result credits GPT-5.6 Sol Ultra: the publicly available commercial model operating in its highest-compute mode.

That shift is meaningful. The unit distance proof involved “a new general-purpose reasoning model, rather than a system trained specifically for mathematics.” The CDCC proof uses the same model consumers access today, via the API’s multi-agent ultra capability.

What Ultra Mode Is

GPT-5.6 Sol Ultra coordinates four agents running in parallel across workstreams by default, recombining results to push beyond what any single reasoning thread sustains. On Terminal-Bench 2.1, Ultra scores 91.9% against Sol’s 88.8%. On BrowseComp, Ultra reaches 92.2% against Sol’s 90.4%.

Extended mathematical proof search is where parallel exploration is structurally advantageous. Human attempts at the CDCC typically pursued one approach — usually inductive arguments on snark reductions, cubic graph decompositions, or flow-based constructions — for years at a time. A system running parallel proof strategies, checking partial results against each other, and recombining has a qualitatively different search profile over the space of possible proofs.

OpenAI has not yet released the model’s chain of thought for the CDCC proof — it did publish an abridged version for the unit distance disproof. When it does, graph theorists will be looking closely at which structural approach succeeded where previous attempts failed.

Verification Status

The proof was posted to HN this morning. External verification has not yet been publicly confirmed. The comparison to the unit distance problem matters here: that proof was checked by a group of external mathematicians before OpenAI announced it. The CDCC is harder and carries more downstream implications. The community will read it carefully.

The mathematics of graph cycle covers is well-defined and machine-verifiable in principle: a formal proof checker could, in theory, validate the argument without human intervention if the proof is written in a suitable form. Whether OpenAI has done this, or submitted to a formal verification system, is not yet known.

If Verified

A confirmed CDCC proof closes a line of inquiry that has run since the early 1970s and removes one of graph theory’s most prominent open problems. It also makes a specific empirical statement about GPT-5.6 Sol Ultra: a commercially deployed reasoning model, using only the parallel-agent capability described in the public API, reached a result that 50 years of specialist human effort did not.

That is a different claim than “AI helped a human solve a problem” or “AI found a counterexample to a conjecture.” It is an affirmative proof of a deep structural result about infinite families of graphs, produced autonomously.

The PDF is public. Graph theorists are reading it now.