OpenAI claims GPT-5.6 Sol Ultra solved a 50-year math conjecture in one hour
The AMW Read
Novelty 2: updates the frontier-model capability baseline with multi-agent scaffolding; significance 2: segment-level impact for foundation models and formal reasoning if validated, but pending verification limits cross-segment reach.
OpenAI claims GPT-5.6 Sol Ultra solved a 50-year math conjecture in one hour
OpenAI researcher Ethan Knight posted on X that the company's latest model, GPT-5.6 Sol Ultra, produced a proof of the Cycle Double Cover Conjecture—a longstanding open problem in graph theory dating to 1973—in under an hour using 64 sub-agents. The system allocated each agent a distinct mathematical approach (algebraic, structural induction, etc.), with an adversarial agent checking for counterexamples and logical errors. The resulting three-page proof reduces the problem to cubic graphs and leverages the 8-flow theorem with GF(3)-based edge labeling. External formal verification using Lean or Coq was not performed, as those libraries lack sufficient graph theory coverage.
Why it matters: This event sits at the intersection of the agentic orchestration pattern and the frontier-model capability debate. OpenAI's architecture—64 parallel sub-agents plus an adversarial verifier—is a canonical instance of the multi-agent scaffolding approach that is emerging as a structural force in foundation model workflows. The proof's conciseness and speed, if validated, would update the baseline for what reasoning-model + agentic scaffolding can achieve in formal mathematics, a domain previously thought resistant to AI-driven discovery. However, the absence of peer review and formal verification means the claim remains provisional, echoing past retracted proofs of this same conjecture—a classic case from the skepticism memory bank.
Ground-evaluation: The graph theory researcher Thomas Bloom (University of Manchester) called the proof "a very nice proof" and praised the model's refusal to abandon failed approaches—an attribute that differs from typical human research behavior. However, he flagged missing citation of the foundational 1983 Bammond, Jackson, and Jaeger paper, a recurring weakness in AI-generated mathematics. Validation by the broader mathematics community over the coming weeks will determine whether this becomes a milestone or a cautionary tale. If confirmed, it would validate the multi-agent scaffold as a repeatable discovery engine; if debunked, it reinforces the need for formal verification infrastructure before such claims can be trusted.
