What is the Cycle Double Cover Conjecture and did GPT-5.6 Sol Ultra really solve it?

The Cycle Double Cover Conjecture, posed by Paul Seymour and George Szekeres in 1979, states that every bridgeless graph has a family of cycles such that every edge belongs to exactly two of them. On July 10, 2026, OpenAI published a proof generated by GPT-5.6 Sol Ultra using 64 subagents in parallel in under an hour, plus the full prompt used. The math community is now reviewing the proof — as of July 12, 2026 it has not been formally peer-reviewed or verified in Lean 4, and expert reaction on Hacker News and social media is split between 'genuine breakthrough' and 'heavily human-scaffolded chain of thought'.

How many subagents did GPT-5.6 Sol Ultra use to prove it?

OpenAI's Ethan Knight posted on July 10, 2026 that GPT-5.6 Sol Ultra used 64 subagents running in parallel and produced the proof in just under one hour. Sol Ultra is the high-effort mode of GPT-5.6 Sol that dynamically orchestrates concurrent subagents across parallel workstreams — priced at ~$12.50 per million input tokens and ~$75 per million output tokens (roughly 2.5x standard Sol pricing).

Is this the same as AI achieving mathematical AGI?

No. The proof was produced with a heavily engineered prompt that fed the model specific strategies to try, and the result still needs formal verification (ideally in Lean 4). It's a milestone in AI-assisted theorem discovery — comparable to DeepMind's AlphaProof solving IMO problems — not evidence that models spontaneously do research-level math. Mistral's Leanstral 1.5 (also announced this month) is the more relevant direction: models that generate proofs already checked in Lean 4.

Can I access GPT-5.6 Sol Ultra to try this myself?

Yes, as of July 9, 2026 GPT-5.6 Sol Ultra is generally available via the OpenAI API, Codex, ChatGPT Work, and ChatGPT Pro. OpenAI published both the proof and the full prompt used, so you can reproduce the run or apply the same subagent-orchestration pattern to your own problems. Budget-limit it explicitly — a single run comparable to the conjecture proof will spend tens to hundreds of dollars in tokens.

Quick Answer

GPT-5.6 Sol Ultra Proved the Cycle Double Cover Conjecture (July 2026)

Published: July 12, 2026

The Short Version

On July 9, 2026 OpenAI made GPT-5.6 Sol Ultra generally available. The next day, engineer Ethan Knight posted on X that Sol Ultra had produced a proof of the Cycle Double Cover Conjecture — one of the most famous open problems in graph theory, posed by Paul Seymour and George Szekeres in 1979 — in under an hour using 64 subagents running in parallel. OpenAI published the proof PDF and the full prompt.

As of July 12, 2026, the proof has not been formally peer-reviewed and has not been verified in Lean 4. The math community is actively reading it.

What the Conjecture Actually Says

For every bridgeless graph G, there exists a family of cycles C₁, …, Cₖ such that every edge of G belongs to exactly two cycles in the family.

“Bridgeless” means no single edge whose removal disconnects the graph.
“Cycle” here means a subgraph where every vertex has even degree (a slightly weaker definition than ‘closed simple walk’).
The conjecture is famous because it’s easy to state, easy to test on small cases, and has resisted proof for 47 years.

What GPT-5.6 Sol Ultra Actually Did

Fact	Detail
Model	GPT-5.6 Sol Ultra
Mode	Ultra (high-effort, subagent orchestration)
Subagents used	64, in parallel
Wall-clock time	Under 1 hour
Announced	July 10, 2026 by OpenAI’s Ethan Knight
Materials released	Full proof PDF + full prompt PDF
Peer review status	None yet as of July 12, 2026
Formal verification (Lean 4)	Not done as of July 12, 2026

What “64 Subagents” Means

Sol Ultra’s headline feature — introduced with GPT-5.6 on June 26, 2026 and GA on July 9 — is the ability for one Sol call to spawn and coordinate up to 64 concurrent subagents. Each subagent explores a different branch of the reasoning tree. A parent agent aggregates and prunes. This is what pushed Sol Ultra to 91.9% on Terminal-Bench 2.1.

For the conjecture proof, the subagents were used to try different proof strategies (nowhere-zero flows, cycle-continuous mappings, ear decompositions, etc.) in parallel, and the parent selected the branch that closed.

Why the Reaction Is Split

The Hacker News thread (item 48863490) and X threads from math researchers converged on three points:

The prompt is very long. Sol Ultra was fed specific proof strategies to try. Skeptics argue this is closer to “AI executes a research plan” than “AI discovers a proof unaided.” Supporters point out that most human collaborations look the same.
No formal verification yet. Until the proof is machine-checked in Lean 4 or Coq, calling it “solved” is premature. Compare: DeepMind’s AlphaProof submissions to IMO 2024 were Lean-verified; Sol Ultra’s proof is not.
The 47-year track record. The Cycle Double Cover Conjecture has famously false-proof-attracted attempts. Historical base rate for “proof of Seymour-Szekeres” being wrong on first pass is high.

How It Compares to Other Recent Math-AI Milestones

Milestone	Date	Model	Verified?
DeepMind AlphaProof — IMO 2024 silver	Jul 2024	AlphaProof + AlphaGeometry 2	Yes (Lean 4)
OpenAI o1 — Putnam 2024	Dec 2024	o1	No (human graded)
Google FunSearch — Cap Set	Dec 2023	Codey + evolutionary search	Yes (proved via constructions)
Mistral Leanstral 1.5	Jul 2026	Leanstral 1.5	Yes (Lean 4 native)
GPT-5.6 Sol Ultra — Cycle Double Cover	Jul 10, 2026	Sol Ultra + 64 subagents	Pending

The pattern to watch: Lean-verified proofs beat prose proofs. If OpenAI or a third party ports Sol Ultra’s proof to Lean 4 without gaps, this becomes the biggest AI-math result of the year. If it doesn’t verify cleanly, it becomes a cautionary tale about long prompts and long proofs.

What Developers Should Take Away

Ultra mode is real and it works on hard problems. 64-subagent orchestration is not a demo — it’s shipping infrastructure. If you have a task that decomposes into parallel branches (large-scale codebase refactor, exhaustive test generation, security audit across many components), Sol Ultra is now the state of the art.
Cost per Ultra call is high. At $12.50/$75 per MTok, a coordinated 64-subagent run on a hard problem burns tokens fast. Budget-limit at the API-key level.
The subagent pattern is copyable. OpenAI published the exact prompt. You can adapt the “spawn N branches, each tries a different strategy, parent prunes” template to your own problem class.

Sources

OpenAI announcement (Ethan Knight on X, Jul 10, 2026): x.com/eknight/status/2075643450196971805
Proof + prompt PDFs (OpenAI CDN): cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98
The Decoder coverage: the-decoder.com/openais-gpt-5-6-sol-ultra-reportedly-solves-a-50-year-old-math-problem-in-under-an-hour
Hacker News discussion: news.ycombinator.com/item?id=48863490