OpenAI’s GPT-5.6 Sol Ultra claims proof of Cycle Double Cover Conjecture; verification pending

OpenAI reports GPT-5.6 Sol Ultra produced a purported proof of a 50‑year‑old math conjecture, independent verification pending

OpenAI reports that GPT‑5.6 “Sol Ultra” produced a document the company says is a complete proof of the Cycle Double Cover Conjecture in under an hour. The claim, reported by Matthias Bastian at The Decoder on July 11, 2026, is striking if verified. Community verification and full reproducibility checks are still pending.

The Cycle Double Cover (CDC) Conjecture, in plain terms, asks whether every bridgeless finite graph has a collection of cycles so that each edge lies in exactly two of those cycles. OpenAI accompanied its announcement with two artifacts cited in the reporting: a proof PDF (cdc_proof.pdf) and a prompt/workflow PDF (cdc_prompt.pdf). Mathematician Thomas Bloom (University of Manchester) has publicly praised the proof’s style while criticizing the write‑up for omitting earlier related work, notably a 1983 paper by Bermond, Jackson, and Jaeger. At the time of the report, broader, independent mathematical confirmation had not yet been completed.

What was reportedly done

Model: GPT‑5.6 Sol Ultra (also referred to as GPT‑5.6 Sol).
Target: Cycle Double Cover Conjecture, a classical open problem dating to the 1970s.
Compute configuration (reported): 64 parallel subagents (OpenAI’s terminology). Implementation details, such as separate model instances, threads, GPUs, orchestration layer, and seeds, have not been fully disclosed in public reporting.
Runtime (reported): just under one hour. The prompt reportedly instructed the system to compute for at least eight hours before giving up, and reporting suggests the run terminated early when a candidate proof was produced. That procedural detail has not been fully clarified.
Prompt constraints (reported): the workflow told the model to assume a proof exists, banned internet searches, forbade partial answers, and included adversarial agents tasked with checking candidate proofs against typical errors.
Artifacts: OpenAI is said to have released a proof PDF and a prompt/workflow PDF as part of the announcement. Independent access to full logs, intermediate agent outputs, and reproducibility metadata remains necessary for thorough evaluation.

Why mathematicians are impressed, and cautious

Thomas Bloom’s reaction captures the community’s mixed view. He praised the proof’s character while flagging an important omission in attribution. As reported in his public comments on X (Twitter):

“a very nice proof, ” noting the solution is “short, elementary, and could have been discovered in the 1980s.”

He also warned about missing citations and the broader pattern he’s seen in AI‑generated mathematics:

“I assume that these previous works were a big influence on the OpenAI proof, and it is a shame that it does not mention them at all […]. This is a frequent issue with AI-generated proofs and papers: they use ideas and proof strategies taken from the literature without proper citation.”

Bloom emphasized why the workflow’s persistence matters, using this example:

“One can imagine trying the natural labelling first, checking the linear algebra, and when that failed shrugging and thinking ‘oh well, I was expecting to fail, guess it can’t be done this easily’ – while the AI does not get discouraged and keeps trying small variations.”

Two clarifications are important. First, “does not get discouraged” is a metaphor: models do not experience psychology, but agentic workflows can keep exploring until termination criteria are met, whereas human researchers may stop pursuing low‑probability avenues. Second, instructing a model to “assume a proof exists” is a methodological choice that can steer search constructively but also raise the risk of producing superficially coherent but incorrect proofs. Adversarial checking and independent review are therefore essential.

A consolidated verification checklist

Before treating the claim as settled, the following verification steps should be completed and publicly documented:

Independent peer review: expert scrutinies and referee reports from graph‑theory specialists.
Independent confirmation/refutation: at least two independent mathematicians reproduce, critique, or correct the argument.
Formalization: machine‑checkable encoding in a proof assistant (Lean/Coq/Isabelle) when the stakes warrant it. Formalization removes interpretive ambiguity but can require substantial effort. Examples include prior large formalizations such as the Feit, Thompson and Kepler projects.
Provenance and reproducibility pack: the minimal set should include the proof PDF, the full prompt text (cdc_prompt.pdf), orchestration logs, intermediate agent outputs, adversarial agent reports, hardware/configuration details, and random seeds or other determinism metadata.

Why this matters for business leaders

Beyond the math, three operational truths matter for organizations using AI for research, design, and optimization.

Parallel, agentic search scales persistence: Splitting a hard search into many parallel workers with adversarial checks can uncover short, human‑readable solutions that a single researcher might abandon. This pattern maps well to combinatorial optimization, automated circuit and layout search, and certain phases of drug or material design where many variants must be evaluated at scale.
Provenance is strategic: If you plan to rely on AI output in IP filings, regulatory submissions, or safety‑critical design, demand reproducibility metadata, citation of prior art, and verifiable checks. Claims without artifacts invite legal and commercial friction.
Not all research translates: Agentic persistence excels where the domain is theory‑mature but search‑hungry. It is less likely to replace open‑ended conceptual research that requires genuinely new theoretical leaps or creative reframing of problems.

Concrete example: an agentic search approach is promising for a combinatorial design problem where existing theory constrains the search space and many small tweaks matter. It’s less suited to inventing new foundational theories where there’s no clear search space to exhaust.

Practical next steps for technical teams (prioritized)

High priority, auditability: Capture and retain orchestration logs, intermediate agent outputs, prompt versions, and compute configurations. These are essential for any downstream validation, IP work, or regulatory use. Cost: modest engineering effort. Critical to trust.
High priority, expert‑in‑the‑loop review: Pair domain experts with AI outputs from day one. Expert triage filters plausible results from artifacts that require deeper proof or formal checks. Cost: expert time, weeks, but far cheaper than unchecked deployment.
Medium priority, formal verification plan: For results that will carry legal, safety, or scientific weight, budget for formalization or third‑party verification, months to years depending on complexity. Expect significant time and specialist effort for full machine‑checked proofs.
Ongoing, governance and citation rules: Establish policies requiring provenance metadata, citation of source literature, and a reproducibility pack before any AI‑generated scientific claim is used externally.

What to watch for next

Release and public availability of the full cdc_proof.pdf and cdc_prompt.pdf, including the exact prompt lines about the eight‑hour directive, the list of adversarial checks, and the subagent definitions.
Preprint or journal submission from OpenAI and any accompanying arXiv entry.
Independent confirmations or critiques from two to three established graph‑theorists, expect community commentary within weeks to months.
Any attempt to formalize the proof in a proof assistant (Lean/Coq), this is a higher bar that, when achieved, substantially raises confidence.

Key questions, short answers

Did GPT‑5.6 Sol Ultra actually prove the Cycle Double Cover Conjecture?

OpenAI reports that it did and published a proof PDF; the report was covered by Matthias Bastian at The Decoder (July 11, 2026). Independent, community‑wide verification and reproducing the run remain pending.
How did the system find the proof so quickly?

OpenAI reports a 64‑way parallel agent workflow, adversarial checkers, and a prompt that enforced persistent search. Exact orchestration details (model instances, hardware, seeds) have not been fully disclosed in public reporting.
Is the proof genuinely new or a recombination of older work?

Thomas Bloom called the proof “short, elementary” and suggested it could have been discovered in the 1980s; he also noted missing citations to earlier work (notably Bermond, Jackson & Jaeger, 1983). The degree to which the argument is novel versus recombinative remains an open question pending side‑by‑side comparisons and citations.
What must happen before the community accepts the result?

Independent peer review, reproduction or refutation by other experts, transparent release of provenance artifacts, and ideally a machine‑checkable formalization for high‑stakes acceptance.

Measured takeaway

Agentic, parallelized AI workflows combined with careful prompt engineering can drastically speed up search‑heavy problems in domains where theory is well developed. That does not yet translate into universal scientific authorship. Provenance, attribution, reproducibility, and formal verification remain the currency of trust. For organizations, the opportunity is real, but so is the obligation to demand artifacts, expert review, and governance before treating headline claims as production‑ready.