July 13, 2026. Sometime late on July 10, OpenAI posted two PDF files and a short announcement: GPT-5.6 Sol Ultra, running 64 subagents in parallel, had produced a complete proof of the Cycle Double Cover Conjecture, a graph theory problem that has been open for roughly 50 years. The run took under an hour. The model had been instructed to work for at least eight. The claim still needs full verification by the mathematical community, and the first expert review raises real questions about novelty. But for anyone running AI agents on business problems, the most valuable artifact here is not the proof. It is the prompt OpenAI published alongside it, which reads less like a chat message and more like a management system for a team of 64 workers that never get tired and never get discouraged.
What happened: the key facts
- The claim. OpenAI announced that GPT-5.6 Sol Ultra produced a complete proof of the Cycle Double Cover Conjecture, posed independently by George Szekeres in 1973 and Paul Seymour in 1979. The conjecture asks whether every bridgeless graph has a set of cycles that covers each edge exactly twice. Decades of partial results existed; no accepted general proof did.
- The setup. The run used 64 subagents working in parallel in ultra mode, the multi-agent setting OpenAI shipped with GPT-5.6 general availability on July 9. It finished in under an hour despite being told to compute for at least eight hours before it was allowed to consider giving up.
- The receipts. OpenAI published both the proof paper, written by GPT-5.6 Sol, and the full prompt that produced it.
- The first expert read. Mathematician Thomas Bloom of the University of Manchester called it "a very nice proof" that is "short, elementary, and could have been discovered in the 1980s." It combines known tools cleverly rather than inventing new theory.
- The criticism. Bloom notes the core ideas trace back at least to a 1983 paper by Bermond, Jackson, and Jaeger, which OpenAI's paper does not cite at all. In his words, this is "a frequent issue with AI-generated proofs and papers: they use ideas and proof strategies taken from the literature without proper citation."
- The caveat that matters. Bloom's assessment is the most detailed public evaluation so far. A full verification by the mathematical community is still pending, and long-open conjectures have a history of attracting claimed proofs that later collapsed, as The Decoder's coverage is careful to flag.
Machine persistence, not machine genius
Bloom's explanation for why humans missed a proof this simple for 50 years is the most useful sentence in the whole story. The key step involved a small, counterintuitive twist. A human mathematician would likely try the natural approach, watch it fail, and move on, thinking "I was expecting to fail, guess it can't be done this easily." The AI, in his words, "does not get discouraged and keeps trying small variations" until one clicks.
He also expects more results like this: problems whose solutions need only existing, well-developed theory plus a lot of patience. That is likely a small share of open problems, and nobody knows in advance which ones. But note what the advantage actually was. Not creativity. Not new theory. Tireless, systematic persistence at scale, which is precisely the quality businesses buy when they automate work with AI.
The prompt is the playbook
Read the published prompt as an operations document and you find six design rules, every one of which transfers to running agents on business problems.
- Ban the easy outs. The prompt tells the model to assume a complete proof exists, forbids it from searching the internet to check whether the conjecture is open, and forbids answering that the problem is unsolved. The model's most likely honest answer was removed as an option, leaving nowhere to go but the work.
- No partial credit. Partial results, reductions to other unproven conjectures, summaries of the state of research, and explanations of why the problem is hard were all rejected as insufficient. Only a complete proof counts as done.
- Keep workers independent. Most of the 64 agents were deliberately kept in the dark about which approach currently looked most promising, forcing independent attempts instead of groupthink around one strategy.
- Make verification adversarial and specific. Dedicated adversarial agents checked every candidate proof against a written list of typical errors, such as closed paths misidentified as cycles. Verification was a staffed role with a checklist, not a final vibe check.
- Budget persistence explicitly. A minimum of eight hours of compute before giving up was allowed. The model finished in one. Most business automations are declared failures after a single unsupervised attempt.
- Publish the receipts. The output had to survive adversarial review and ship as a complete, checkable artifact. That is what made outside scrutiny, including Bloom's citation critique, possible at all.
What this does not prove
Honest framing matters here. First, the proof has not yet been verified by the broader mathematical community, and that process exists because impressive-looking proofs fail it regularly. Second, Bloom doubts the AI invented its core strategy from nothing, given that a reasoning model's first instinct is to read every related paper. Recombination of known ideas at superhuman speed is still the most defensible description of what happened. Third, labs attack many open problems at once and publicize the wins, so you are seeing the survivors, not the distribution. None of that shrinks the operational lesson, because recombination plus persistence is exactly what most business work needs.
What it means for operators
The gap between a demo agent and this result is not model access. Ultra mode shipped to ChatGPT Work Pro and Enterprise plans and Codex Plus and above on GA day, and the Responses API has a multi-agent beta. The gap is the harness: the rules about what counts as done, who checks the work, and how long the system persists. That harness is a design skill, the same one that separates the agents that need constant correction from the ones that run unattended.
A starter recipe for applying the same pattern to a business problem, whether that is reconciling a messy product catalog, hunting a revenue leak across systems, or exhaustively testing a pricing page. Pick work where the output is machine checkable, so "done" is a test that passes, a ledger that balances to zero, or a form that submits, not "looks good." Write the ban list: no summaries of the problem, no partial answers, no "this needs a human decision" until a defined effort budget is spent. Split attempts across independent runs rather than one long chat, and keep them blind to each other. Then staff verification as its own agent with a checklist of your known failure modes. This is the architecture work an AI engineer does when a workflow has to survive without supervision, and it is the difference between paying for 64 agents of theater and 64 agents of throughput. If the whole pattern is new to your business, start smaller: our team builds these verification-first swarms as part of AI automation projects, typically beginning with one workflow where failure is cheap and success is measurable.
The 50-year problem fell to patience, independence, and adversarial checking. Your backlog will fall to the same three things long before it requires a frontier breakthrough.
Frequently Asked Questions
It is a graph theory problem posed independently by George Szekeres in 1973 and Paul Seymour in 1979. It asks whether every bridgeless graph has a collection of cycles that covers each edge exactly twice. Many partial results existed for special cases, but no generally accepted proof until OpenAI's July 10, 2026 claim.
Not fully. Mathematician Thomas Bloom of the University of Manchester published the most detailed early assessment, calling it a very nice, elementary proof, but a complete verification by the mathematical community is still pending. Long-open conjectures have a history of claimed proofs that later collapsed, so treat it as a strong claim under review.
Ultra is the multi-agent setting OpenAI shipped with GPT-5.6 general availability on July 9, 2026. It runs parallel subagents on one task, with 4 agents as the default setting in products and larger configurations like the 64-agent proof run available via the Responses API multi-agent beta. In products it is available on ChatGPT Work Pro and Enterprise plans and Codex Plus and above.
The published prompt banned the easy outs (it forbade answering that the problem is unsolved and forbade checking the internet), rejected partial results, kept most of the 64 agents blind to the leading approach so attempts stayed independent, ran adversarial verifier agents with a checklist of typical proof errors, and set a minimum effort budget of eight hours. That structure, not a magic phrase, produced the result.
No. The expert consensus so far is that the model recombined known ideas from the literature with superhuman persistence, and it did not cite that prior work, which Bloom criticized. Experts still define the problems, judge the outputs, and catch what the AI misses. The realistic takeaway is that patient, well-harnessed AI can clear work that stalled for lack of persistence, not for lack of genius.
Pick a workflow with a machine-checkable definition of done (a reconciliation that balances, a test suite that passes, a form that submits). Ban partial answers until an effort budget is spent, run several independent attempts in parallel, and add a separate verifier agent with a checklist of your known failure modes. An AI automation partner can build this harness around your existing tools; the pattern works with any frontier model.