# OpenAI says GPT-5.6 Sol Ultra produced a proof of a 50-year graph theory problem > Source: > Published: 2026-07-10 20:09:46+00:00 [Ethan Knight (@](https://x.com/__eknight__)**eknight**) said in a [two-post thread on X](https://x.com/__eknight__/status/2075643450196971805) on Friday that GPT-5.6 Sol Ultra produced a proof of the Cycle Double Cover Conjecture, a graph theory problem that standard references still list as open after roughly five decades. The claim landed one day after [OpenAI made GPT-5.6 generally available](https://openai.com/index/gpt-5-6/) on July 9th. Knight said the proof was generated with 64 subagents in just under one hour. OpenAI published both the [three-page proof](https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_proof.pdf) and the [prompt used to produce it](https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_prompt.pdf) on its CDN. The careful word is produced, not settled. The proof is a candidate proof posted by OpenAI, with no journal acceptance, arXiv posting, or visible public referee process attached to the document. That distinction matters more in mathematics than in software demos. A proof of this conjecture would need scrutiny from graph theorists, especially because the history of long-running conjectures includes plausible-looking proofs that fail under expert review. ### What OpenAI is claiming The Cycle Double Cover Conjecture asks whether every bridgeless graph has a collection of cycles such that every edge appears exactly twice. [Wolfram MathWorld's page](https://mathworld.wolfram.com/CycleDoubleCoverConjecture.html), last updated July 10th, described the conjecture as open and independently formulated by Szekeres in 1973 and Seymour in 1979. A [2017 Journal of Combinatorics paper](https://www.intlpress.com/site/pub/files/_fulltext/journals/joc/2017/0008/0002/JOC-2017-0008-0002-a006.pdf) called it one of the famous open problems in graph theory and summarized partial progress on related cycle-cover variants. OpenAI's PDF states a theorem that every finite bridgeless undirected graph has a cycle double cover. The note says the proof reduces the problem to loopless cubic graphs, uses a nowhere-zero flow over the group F_2^3, and then converts that flow into labels by two-element sets so that each group element appears zero or twice at each vertex. The last step is presented as a linear algebra argument establishing a compatibility condition across edges. The proof document also contains a direct AI-use statement: the proof is attributed to GPT 5.6 Sol Ultra, while the writeup is attributed to Codex with GPT 5.6 Sol. That makes the artifact unusually explicit for a frontier-model claim. OpenAI is not merely saying the model helped a human mathematician search for a route; the document presents the model as the source of the proof. The prompt is equally revealing. It instructed the model to resolve the conjecture completely and said partial progress would not count unless it implied the full resolution. It also told the system to use up to 64 concurrent agents and to keep adversarial agents checking for common failure modes: exact-two multiplicity, parallel-edge 2-cycles, disconnected graphs, bridges introduced by reductions, and circular reliance on an equivalent version of the conjecture. ### The product play behind the math claim OpenAI's timing is deliberate. GPT-5.6 is being sold around a scaling argument for agentic work: spend more compute on harder tasks, get better work back faster. The official launch page says GPT-5.6 spans Sol, Terra, and Luna, and describes ultra as the highest-capability setting, coordinating multiple agents across parallel workstreams. OpenAI says the default ultra setup coordinates four agents, while developers can build multi-agent workflows through the Responses API. Knight's 64-subagent proof run is a sharper demonstration than another benchmark table. Benchmarks tell buyers where a model sits on a leaderboard. A candidate proof of a named open problem tells buyers what OpenAI wants them to believe about the ceiling of its agent orchestration: that hard work can be decomposed, attacked in parallel, audited internally, and synthesized into a compact final artifact. That is also where the commercial tension sits. OpenAI says ultra trades higher token use for stronger results and faster time-to-result. The GPT-5.6 launch page lists Sol API pricing at $5 per 1M input tokens and $30 per 1M output tokens, with Terra and Luna priced lower. If ultra becomes the mode users reach for when ordinary prompting fails, OpenAI gets a path to sell more compute without framing the upsell as a simple model-size upgrade. The conjecture is a useful target for that argument because it is legible outside the narrow AI benchmark circuit. It has a crisp statement, a long history, and a body of partial results. It is also unforgiving. A single hidden circular step or mishandled reduction can collapse the proof. ### The next test is external checking The most important unanswered question is whether graph theorists accept the argument. OpenAI's PDF is short enough that specialists can inspect it quickly, and the prompt gives enough detail to reconstruct what the model was asked to do. That openness is useful, but it is not validation. There is precedent for caution. Searchable mathematical archives include earlier manuscripts claiming proofs of the Cycle Double Cover Conjecture, while reference pages and recent literature have continued to describe the problem as open. That gap is common in mathematics: a claimed proof becomes a theorem only after the community finds the argument complete. If the proof survives, OpenAI will have a stronger story than a model release. It will have evidence that multi-agent test-time compute can generate publishable mathematical work on a problem outside benchmark design. If it fails, the episode still matters because it shows how frontier labs are going to market their most expensive modes: with public artifacts that force outside experts to distinguish real advances from polished overreach.