Event timeline: announced by OpenAI on 2026-07-10, paper posted 2026-07-11, indexed by aihot 2026-07-12. Sources: IT之家 roundup and Hacker News. aihot score: 84. Links: aihot, original article.
What Happened
On July 10, 2026, OpenAI announced that GPT-5.6 Sol Ultra generated a complete proof of the Cycle Double Cover Conjecture in under one hour. The conjecture, posed independently by George Szekeres (1973) and Paul Seymour (1979), asks: does every bridgeless graph admit a collection of cycles such that every edge appears in exactly two of them?
Key facts released by OpenAI:
- The model orchestrated 64 parallel subagents plus several Adversarial Agents.
- The budgeted compute time was 8 hours; actual usage was about 1 hour.
- The proof was produced fully independently by AI; OpenAI published both the proof PDF and the prompts on its CDN.
- Cost at official Sol pricing: roughly $275-485 (~1,867-3,293 RMB); up to $13,000 on the Cerebras platform.
- Up to 64 concurrent parallel subagents with dynamically managed workloads;
- Early-stage diversity of research routes (different mathematical representations, algebraic approaches, structural induction);
- Dedicated adversarial agents hunting for holes, edge cases, and errors;
- No web search allowed;
- Refusal of special-case-only or incomplete proofs;
- Mandatory adversarial verification against common mathematical errors.
Key technical steps of the proof:
1. Reduce the conjecture to the cubic graph case. 2. Apply the 8-flow theorem. 3. Construct an edge labeling via linear algebra over GF(3). 4. Show every edge belongs to exactly two cycles.
OpenAI researcher Ethan Knight announced on X: "Yesterday, we made GPT-5.6 Sol Ultra available to all users. Today, we're sharing that it proved the 50-year-old Cycle Double Cover Conjecture in under an hour using 64 subagents."
Mathematician Thomas Bloom (University of Manchester) was among the first to comment: "This is a very beautiful proof" — concise, elementary, using methods that could have produced it in the 1980s had anyone thought to try. He highlighted AI's core advantage: "far beyond human patience in trying small variations."
Deep Analysis
The author argues this matters more than DeepMind's 2023 Cap Set result or AlphaProof's IMO performance: it is the first demonstrated case of an LLM independently proving a problem on Wikipedia's list of unsolved mathematical problems — independent rather than AI-assisted, since prior milestones involved human iteration or prompt engineering.
OpenAI's published prompt spec is unusually strict:
1. Parallel + diverse exploration via 64 subagents — extending multi-agent software engineering practice into mathematical proof exploration; described as the first public spec of 2026's LLM reasoning-engineering paradigm. 2. Adversarial agents as a separate role — splitting "solve" and "check" into distinct agents; the first public industrialization of Adversarial Verification. 3. Hard constraints in the system prompt (no internet, no partial proofs, adversarial validation) — a reusable template for downstream AI-math workflows. 4. 8-hour budget, 1-hour usage — implying OpenAI has a well-calibrated engineering estimate of its model's reasoning ceiling (budget set at roughly 4-8× the estimated median).
The deeper takeaway: Bloom's remark that 1980s tools sufficed suggests the 50-year stall was not due to mathematical difficulty but to a lack of idea diversity plus verification patience. AI's core innovation here is not new mathematics but exhaustive trial within combinations of classical tools.
Why It Matters
Four simultaneous boundary expansions:
1. From AI-assisted solving to independent proofs of listed unsolved problems — an "X moment" for LLM reasoning, likely to boost AI-inference investment in late 2026. 2. From single-model inference to multi-agent reasoning engineering — the published spec serves as an entry template for serious AI math/science work; platforms like OpenRouter, Groq, and Cerebras may offer "multi-agent reasoning as a service." 3. From mathematical discovery to industrialized discovery — expect many "old problem + old tool combinations → AI retry → sudden breakthrough" cases. 4. First public real cost of research-grade inference — $275-485 for a 50-year problem, comparable to an AWS Lambda bill; a genuine starting point for democratized AI for Science.
Risks and Open Questions
1. No peer review, no citations, no formal verification. OpenAI disclosed these details itself. Past claimed proofs of the conjecture on arXiv were refuted; Bloom noted the proof does not cite the classic Bermond/Jackson/Jaeger 1983 paper. The math community is expected to complete line-by-line review within days to weeks — a substantive error would be a blow to OpenAI's technical credibility. 2. Is AI's "patience with small variations" really superior? Bloom observed that human mathematicians typically try one natural approach and abandon it on failure. A 64-person team each spending a week approximates the AI's compute budget; AI's edge may be simulating team-level research rather than surpassing any individual. 3. Cost comparisons are not uniform. The same task could cost up to $13,000 on Cerebras; for typical institutions the realistic range is $1,000-10,000 — not zero-barrier. 4. A victory of tool combination, not methodology. No new mathematical theory was created. The hardest problems (P vs NP, the Riemann Hypothesis) will still require genuinely new ideas, not just AI patience.
The most consequential detail may be that OpenAI published the prompts. With this spec public, Anysphere, Cognition, Anthropic, xAI, DeepMind, Zhipu AI, and Moonshot AI are expected to follow quickly — multi-agent reasoning engineering now has its entry template.