GPT-5.6 Sol Ultra Proves 50-Year Math Conjecture Today

OpenAI announced on July 10, 2026, that its GPT-5.6 Sol Ultra model proved the Cycle Double Cover Conjecture, a 50-year-old math problem, using 64 parallel AI subagents in under an hour. The proof, published as a PDF alongside the prompt, uses elementary graph theory and awaits peer review. The achievement highlights both the potential of multi-agent AI systems and the challenge of verifying opaque AI-generated proofs.

On July 10, 2026 — today — OpenAI announced that GPT-5.6 Sol Ultra proved the Cycle Double Cover Conjecture using 64 parallel AI subagents in under one hour. The conjecture, posed independently by Szekeres in 1973 and Seymour in 1979, had resisted proof for roughly 50 years. OpenAI’s Ethan Knight published the announcement on X alongside a public PDF of the proof and the full prompt used to generate it. The Hacker News thread hit 179 upvotes and 165 comments within hours. GPT-5.6 Sol reached general availability just yesterday. The same Ultra mode that cracked this problem is already in the OpenAI API. What the Cycle Double Cover Conjecture Actually Is The Cycle Double Cover CDC Conjecture asks a deceptively simple question: does every bridgeless graph have a collection of cycles such that each edge appears in exactly two of those cycles? A bridgeless graph is one where removing any single edge doesn’t disconnect it. The conjecture sounds manageable until you start trying to prove it for “snarks” — bridgeless cubic graphs that can’t be cleanly partitioned into three perfect matchings. Snarks exist with arbitrarily high girth, which means any finite reduction strategy eventually fails. That’s the wall that stopped human mathematicians for 50 years. GPT-5.6 Sol Ultra apparently found a way around it. The proof runs approximately three pages, uses elementary graph theory rather than novel frameworks, and was described by a graph theory researcher in the HN thread as “prose reading like an old paper — straightforward theorems with direct proofs.” OpenAI published the full proof as a PDF https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc proof.pdf alongside the exact prompt used to generate it. How 64 GPT-5.6 Subagents Pulled This Off GPT-5.6 Sol Ultra doesn’t run a single reasoning chain. It decomposes problems and spawns parallel subagents that are, in OpenAI’s words, “trained to cooperate and allowed to communicate with each other during a task.” The default configuration is four cooperative subagents. For the CDC proof, OpenAI scaled this to 64. The result landed in under an hour. The trade-off is real and worth naming: there is no inspectable transcript. You get a single opaque result with no record of how the subagents disagreed, explored dead ends, or resolved conflicts. HN commenters noted the coordination “does not obviously fit standard LLM architecture,” suggesting novel inference-time mechanisms OpenAI hasn’t published. Developers who want similar workflows can access the multi-agent beta in the Responses API https://openai.com/index/gpt-5-6/ today. However, the irony is hard to miss: the same system that can’t show its reasoning is the one being asked to generate mathematical proofs that humans must verify by hand. Related: GPT-5.6 Sol, Terra, and Luna: Developer Routing Guide The Community Is Divided — Appropriately The Hacker News discussion https://news.ycombinator.com/item?id=48863490 captures two camps cleanly. The first is genuinely excited: a 50-year-old problem, three pages, elementary methods, under an hour. A graph theory researcher in the thread called the prose “like an old paper” — implying the solution exploited existing theory cleverly rather than stumbling into something plausible-sounding. The second camp wants peer review before celebrating. One commenter put it directly: “Without review from professional mathematicians, you could easily generate thousands of plausible-looking PDFs.” The counter-argument is that at three pages using elementary methods, expert review is straightforward — far easier than verifying a typical 40-page proof. What’s missing is Lean formalization: graph theory lacks the mature formal libraries that would let a proof checker confirm this automatically. That’s the honest gap. OpenAI announced a result; the math community still needs to verify it. The prompt architecture is worth noting too. Developers examining the released prompt found that only about one-fifth of it addresses the actual mathematical problem. The remaining four-fifths optimize the model’s behavior — task persistence, self-correction, exploration strategy. The math almost looks like the smaller part of the job. What This Changes for Developers The practical takeaway is sharper than “AI solved a math problem.” GPT-5.6 Sol Ultra scored 91.9% on Terminal-Bench 2.1, putting it 3.9 points ahead of the next competitor Claude Mythos 5 and GPT-5.5 at 88.0% . HN commenters estimate the CDC proof run cost between $275 and $13,000 — accessible to any well-funded team. At Sol’s pricing $5 input / $30 output per million tokens , attacking a hard structured problem no longer requires a research institution. The question that matters now: what other hard, structured problems — in biology, cybersecurity, code analysis, logistics — fall to 64 cooperative subagents working in parallel? The Cycle Double Cover Conjecture https://en.wikipedia.org/wiki/Cycle double cover was one data point. The multi-agent beta is open. If the CDC proof holds up, someone is about to start a list. Key Takeaways - GPT-5.6 Sol Ultra reportedly proved the 50-year-old Cycle Double Cover Conjecture today using 64 cooperative subagents in under one hour — but “announced” is not the same as “verified” - The proof is approximately three pages of elementary graph theory, making peer review accessible — but Lean formalization is still missing - Ultra mode’s cooperative subagents communicate in real time but produce no inspectable transcript, raising a genuine question about verifiability for mathematical work - The same multi-agent capability is in the OpenAI Responses API now — the cost floor for attacking hard structured problems has dropped significantly - Sol Ultra scored 91.9% on Terminal-Bench 2.1, 3.9 points ahead of the next competitor — the benchmark gap is real