OpenAI GPT-5.6 Sol Ultra Proves 50 Year Graph Theory Conjecture in One Hour

OpenAI GPT-5.6 Sol Ultra solved a graph theory problem that had stood for more than fifty years. The model produced a complete proof of the cycle double cover conjecture in roughly one hour. The result came from a system that ran 64 parallel sub agents together with opposing agents inside an eight hour window. OpenAI published the proof and the prompt set as a PDF.

The cycle double cover conjecture dates to work by George Szekeres and Paul Seymour in the 1970s. It asks whether every bridgeless graph contains a collection of cycles that together cover every edge exactly twice. No human proof has appeared in the intervening decades. Wikipedia still lists the statement among major unsolved problems in mathematics.

Proof Appeared After One Hour of Agent Coordination

The model did not work alone. It launched 64 separate reasoning paths at the same time. Some paths acted as critics that tried to break partial arguments while others attempted to strengthen them. The process finished inside the first hour of the allotted eight hours. OpenAI released the resulting document and the exact prompt sequence without further editing.

The released PDF contains the proof text and the full set of instructions given to the model. No Lean code or other formal verification accompanies the file. The announcement states that the proof remains unexamined by human experts.

Stakes Rise for Every Group Building Automated Reasoning Tools

A working proof would mark the first time an LLM independently settled an item on the Wikipedia list of unsolved mathematics problems. Several research groups have spent years on narrower problems with formal systems such as Lean. Those projects required teams of mathematicians and months of effort. A one hour run therefore forces every lab to compare its timeline against the new result.

Companies that sell AI coding assistants now face direct questions about verification speed. Investors who backed formal methods startups must decide whether the new route reduces or increases the value of their holdings. Academic departments that train students in graph theory must consider how curricula will change once machine generated outlines become routine.

OpenAI Internal Agent Design Faces Direct Comparison With Lean Based Efforts

OpenAI chose a multi agent architecture that keeps all reasoning inside natural language. In contrast, groups at Microsoft Research and the University of Cambridge have focused on exporting statements into Lean and checking them mechanically. The cycle double cover case therefore pits two distinct technical bets against each other. One side relies on rapid iteration inside language. The other side insists on machine checked steps before any claim is accepted.

The multi agent result arrived faster, yet it offers no mechanical certificate. The Lean route moves slower but produces an artifact that can be checked by any computer in seconds. Funding agencies will watch which approach secures the next major theorem first.

Verification Gap Remains the Central Uncertainty

The proof has not entered formal review. No journal has accepted it. No independent team has rebuilt the argument inside a proof assistant. OpenAI itself notes these limits in the release notes. Until those steps occur, the result stays in the category of notable but unconfirmed output.

Skeptics point out that earlier language model proofs contained subtle gaps that only surfaced after human inspection. Supporters reply that the speed of generation at least allows faster human review cycles. Both positions rest on the same fact: the current document carries no formal guarantee.

Three Signals Will Show Whether the Result Holds

First, any announcement that a named mathematician has checked the argument line by line will indicate progress toward acceptance. Second, an attempt to encode the proof in Lean or another assistant will reveal whether the language steps translate without contradiction. Third, a replication run by an external group using a different model will test whether the outcome depends on OpenAI specific scaffolding.

Each of these checkpoints can occur within the next three months. Any one of them would either strengthen or weaken the original claim in measurable ways.