GPT-5.6 Sol Ultra Proves a 50-Year-Old Graph Theory Conjecture in Under an Hour — A Paradigm Shift via "64 Subagents + Mathematical Patience"
> Event timeline: 2026-07-10 (OpenAI announcement) / 2026-07-11 (paper posted) / 2026-07-12 (indexed) > Sources: IT之家 summary + Hacker News > aihot score: 84 (highest) > aihot link: https://aihot.virxact.com/items/cmrfdbe4g016aihxb8btg5nu0 > Original link: https://www.ithome.com/0/975/646.htm
What Happened
On July 10, 2026, OpenAI announced that its GPT-5.6 Sol Ultra model produced a complete proof of the Cycle Double Cover Conjecture in graph theory in under one hour.
The conjecture was independently proposed by mathematician George Szekeres (1973) and Paul Seymour (1979), and had remained open for over 50 years. Its standard statement: does every bridgeless graph admit a collection of cycles such that every edge appears in exactly two of them?
Key figures released by OpenAI:
- The model invoked 64 parallel subagents plus several adversarial agents;
- The originally reserved compute budget was 8 hours; actual usage was about 1 hour;
- The entire proof was completed independently by AI; OpenAI published both the proof PDF and the prompts on its company CDN;
- Compute cost at OpenAI's official Sol pricing: roughly $275–485 USD (~1,867–3,293 RMB); running the same task on the Cerebras platform could cost up to $13,000.
- Up to 64 parallel subagents running simultaneously;
- Dynamic management of each agent's workload;
- Maintaining diversity of research directions early on, with different agents trying different mathematical representations, algebraic approaches, and structural induction;
- Dedicated "Adversarial Agents" tasked with hunting for holes, edge cases, and potential errors;
- No internet access allowed;
- Rejection of special-case-only or incomplete proofs;
- Mandatory adversarial verification checking for common mathematical errors.
Key technical steps of the proof:
1. Reduction of the original conjecture to the cubic graph case; 2. Application of the 8-flow theorem; 3. Construction of edge labeling via linear algebra over GF(3) (the finite field of three elements); 4. Final proof that every edge belongs to exactly two cycles.
OpenAI researcher Ethan Knight announced the result on X: "Yesterday we made GPT-5.6 Sol Ultra available to all users. Today, we're sharing that it used 64 subagents to prove the 50-year-old Cycle Double Cover Conjecture in under an hour."
Mathematician Thomas Bloom (University of Manchester, UK) was among the first to comment publicly: "This is a very beautiful proof." "Elegant, elementary, using methods that aren't sophisticated — if someone had thought of it, this proof could have been completed in the 1980s." He also highlighted AI's core advantage: "far superhuman patience in repeatedly trying small variations."
Deep Analysis
The engineering significance, in my view, exceeds DeepMind's 2023 Cap Set Problem solution and AlphaProof's IMO results — it is the first demonstrated case of an LLM independently proving a problem on Wikipedia's list of unsolved mathematical problems.
Why "independent" rather than "assisted"? DeepMind's Cap Set work was an iterative human-AI collaboration, and AlphaProof's IMO results involved "AI solving + human prompt engineering." OpenAI's published prompt spec is strikingly hardcore:
Marker 1: Parallel + diverse exploration with 64 subagents. This is the "multi-agent in parallel" pattern long discussed in AI coding circles, but applied to mathematical proof — transplanting multi-agent collaboration from the code IDE to the mathematician's multi-track exploration. Arguably the first public spec of the 2026 LLM reasoning engineering paradigm.
Marker 2: Adversarial agents as a separate role. Traditional AI reasoning is a two-step "solve → check." OpenAI split "checking" into dedicated agents whose sole job is finding flaws — the first public instance of industrialized adversarial verification.
Marker 3: Three hard constraints in the system prompt — no web search, no special-case-only proofs, mandatory adversarial verification. This is OpenAI's answer to "what counts as a good mathematical proof": diverse exploration + counterexample elimination + self-consistent verification. This spec is fully reusable by any downstream AI-math workflow.
Marker 4: 8-hour budget, finished in 1 hour. Generous budgets are engineering normal for AI reasoning, but a 1/8 utilization rate implies OpenAI has a fairly accurate engineering estimate of its model's reasoning ceiling — the 8 hours was not arbitrary, but roughly 4–8× the median estimated from similar past tasks.
A deeper judgment: Bloom's remark that "this could have been done in the 1980s if someone had thought of it" is the key insight — the reason 50 years passed without a proof was not "mathematical difficulty" but "diversity of approaches + verification patience." The core innovation of AI reasoning here is not "new mathematical ideas" but "exhaustive trial-and-error within combinations of existing classical tools." This is the underlying truth of the AI reasoning engineering paradigm.
Why It Matters
This event pushes boundaries on four levels simultaneously:
Level 1: From "AI-assisted solving" to "AI independently proving Wikipedia-listed open problems." This is the "X moment" for LLM reasoning. Mathematicians took 50 years; AI took 1 hour — the combined media and technical effect will directly fuel investment in the AI reasoning space in H2 2026.
Level 2: From "single-model reasoning" to "multi-agent reasoning engineering" as the next standard paradigm. The spec OpenAI published — 64 subagents + adversarial agents + rejection of special-case-only proofs — can serve as the starter template for any team pursuing "serious mathematical/scientific AI reasoning." Platforms like OpenRouter, Groq, and Cerebras will likely launch "multi-agent reasoning as a service" offerings within the year.
Level 3: From "mathematical discovery" to "industrialized mathematical discovery." A 50-year-old conjecture cracked by "combinations of classical mathematical tools" means we can expect many "old problem + old tool combinations → AI retry → suddenly solved" episodes in mathematics. This is not a threat to mathematicians but the dawn of industrialized mathematical discovery — AI makes the pace of mathematical discovery controllable.
Level 4: First public disclosure of real "reasoning pricing." $275–485 at official Sol pricing means the cost of proving a "50-year problem" is comparable to an AWS Lambda bill. OpenAI has pulled "research-grade" AI reasoning costs into the low four figures of dollars — the true starting point of democratized AI for Science.
Risks and Open Questions
But several uncertainties deserve caution:
Risk 1: No peer review, no citations, no formal verification. These are details OpenAI itself disclosed. Historically, multiple "proofs" of the Cycle Double Cover Conjecture on arXiv have been refuted. Bloom called it "beautiful" while admitting it "doesn't cite the classic 1983 Bermond/Jackson/Jaeger paper." The math community expects a line-by-line review within days to weeks. A substantive error would be a blow to OpenAI's technical credibility.
Risk 2: Is AI's "persistent trial of small variations" in math truly superior to mathematicians'? Bloom noted that "human mathematicians typically try one natural approach and give up if it fails" — a fair characterization, but possibly just a matter of insufficient attempts. A 64-person team of mathematicians each spending a week on different ideas roughly equals 64 subagents × 1 week of AI reasoning budget. AI's real advantage may not be "surpassing an individual mathematician" but "simulating team-level mathematical research."
Risk 3: The cost comparison isn't apples-to-apples. $275–485 is OpenAI's official price; the same task on Cerebras could reach $13,000 — meaning AI reasoning cost depends heavily on where it runs. For ordinary research institutions, schools, or math hobbyists, "running a 50-year conjecture" on Claude Sonnet 5 still costs in the $1,000–10,000 range. It is not "zero-barrier."
Risk 4: This is a victory of "tool combination," not "methodology." Bloom made it clear: the toolset existed in the 1980s. AI created no new mathematical theory — it just permuted existing tools. The truly hardest "Millennium Problems" (P vs NP, the Riemann Hypothesis) will still require genuinely new ideas, not AI "patience."
Finally, the most memorable aspect of this event is not "AI proves a 50-year conjecture in under an hour," but that OpenAI published the prompt as well. With this spec public, Anysphere, Cognition, Anthropic, xAI, DeepMind, Zhipu AI, and Moonshot AI will all follow quickly. Multi-agent reasoning engineering now has its starter template.
*Note: This post describes a claim announced by OpenAI that has not yet been peer-reviewed.*