OpenAI's GPT-5.6 Sol Claims Proof of 50-Year-Old Math Conjecture Using 64 AI Subagents

Less than 24 hours after GPT-5.6 Sol reached general availability, OpenAI published what it described as a proof of the Cycle Double Cover Conjecture — a problem that has resisted the efforts of professional mathematicians for more than 50 years. The announcement, posted on July 10 by OpenAI research lead Ethan Knight alongside a public PDF and the full prompt used to generate it, sent ripples through both the AI and mathematics communities.

What Is the Cycle Double Cover Conjecture?

The conjecture was posed independently by George Szekeres in 1973 and Paul Seymour in 1979. It belongs to a branch of mathematics called graph theory and states the following: for any bridgeless graph — meaning a graph in which no single edge, if removed, would disconnect the graph into two separate pieces — there exists a collection of cycles that together cover every edge of the graph exactly twice.

Intuitive as it sounds, the conjecture has never been proven in full generality. Partial results accumulated over five decades: mathematicians established it for specific families of graphs, derived increasingly tight constraints, and proposed many partial approaches, but a complete proof remained elusive. The problem sits in the same rarefied tier as Goldbach’s conjecture and Collatz’s conjecture — famous enough to be named, simple enough to state, resistant enough to define careers.

How Sol Approached the Problem

OpenAI’s system deployed GPT-5.6 Sol Ultra — the highest-capability tier of the newly launched GPT-5.6 family — in an agentic configuration involving 64 parallel subagents. Rather than a single chain-of-thought pass, the subagents divided the problem, explored different proof strategies concurrently, and synthesized results. The entire run completed in under one hour.

The resulting proof leverages the 8-flow theorem, a deep result in graph theory proven by Jaeger in 1979, and extends it using linear algebra over GF(3), a finite field with exactly three elements. This mathematical structure allows certain graph properties to be encoded algebraically, reducing the cycle cover problem to a question about the solvability of linear systems — a domain where AI reasoning has shown increasing reliability in formal contexts.

OpenAI published the prompt used to generate the proof alongside the PDF, a degree of transparency unusual for mathematical claims of this magnitude. “We’re publishing the prompt and the full output because we want this to be checkable,” Knight wrote on X. “We believe the proof is correct. We also believe it should be verified.”

A Claim, Not Yet a Theorem

The mathematical community’s response has been cautious but genuinely engaged. Discussions on Hacker News and MathOverflow within hours of publication focused on a central tension: machine-verified logical consistency is not the same as mathematical understanding.

“A computer can confirm that every step follows from the last,” wrote one researcher with a background in formal methods. “What we don’t yet have is a sense of why the proof works — what insight it encodes. That matters for building on it.”

The distinction is important. Several previous claimed proofs of the Cycle Double Cover Conjecture have appeared over the years, a handful posted to arXiv, and each was eventually found to contain gaps or withdrawn. None involved AI reasoning, but the pattern underscores why the community is right to proceed carefully. A CDN-hosted PDF, however transparent the publication process, is not a peer-reviewed paper.

What makes this case structurally different from informal AI-generated mathematics is the use of formal verification tooling. The subagent pipeline reportedly used Lean 4-style proof checking at intermediate steps, meaning individual logical transitions were machine-verified rather than merely plausible. This is the same approach Mistral’s Leanstral 1.5 — also released this month — takes to formal software verification, and it provides a meaningfully higher standard of assurance than natural-language reasoning alone.

Still, even a formally verified proof can contain a subtler error: a flawed axiom, a misapplication of a cited theorem, or a gap in the reduction from the original problem to the formalized version. These are exactly the kinds of errors that require expert human scrutiny.

The Broader Context: AI and Mathematics

This announcement arrives at an inflection point for AI-assisted mathematical research. Earlier this year, DeepMind’s AlphaProof system solved several International Mathematical Olympiad problems at a gold-medal level. In May, a collaboration between MIT researchers and an AI agent found a novel short proof of a longstanding combinatorics result that had been known but considered irreducibly complex. The Cycle Double Cover claim, if verified, would represent a significant step beyond those achievements in both difficulty and mathematical depth.

OpenAI has indicated it plans to submit the proof to a peer-reviewed mathematics journal, with independent verification expected to take several weeks. The company has also reached out directly to researchers who have published extensively on the conjecture, offering access to the full agentic trace — the complete log of subagent interactions — for scrutiny.

For researchers in formal mathematics, the proof represents a novel test case: can AI reasoning, scaffolded with formal verification tooling and massive parallel computation, tackle open problems that require not just pattern recognition but genuine mathematical creativity? The answer, over the next few weeks, will come from mathematicians, not from OpenAI.

What It Means for the Field

Whether or not this particular proof holds up, the announcement signals a qualitative shift in what AI systems are being pointed at. The same week GPT-5.6 Sol began general availability, the model is being tested against problems that define the frontier of human mathematical knowledge. That alone marks a departure from the benchmark-driven evaluation cycles that have characterized AI progress for most of the past decade.

For enterprise and developer audiences, the more immediately practical takeaway may be in the method: 64 subagents, operating in parallel, each handling a bounded subproblem, collectively solving something no individual agent — and no human — had managed alone. That architecture is already available through the OpenAI API. The Cycle Double Cover proof may be the most dramatic demonstration of what agentic, parallel AI computation can accomplish, but it is unlikely to be the last.

The mathematics community will have its answer soon. In the meantime, one of the oldest unsolved problems in graph theory has at least one more claimant — and this time, it runs at 750 tokens per second.