On July 14, 2026, OpenAI announced a remarkable breakthrough in mathematics achieved by its latest large language model, GPT-5.6 Sol. This artificial intelligence successfully solved a longstanding mathematical problem known as the "cycle double cover conjecture," a puzzle that has confounded mathematicians for more than fifty years. The solution was unveiled to coincide with the model's full public release, highlighting the growing role of AI in advancing pure mathematical research.
The cycle double cover conjecture is a problem in graph theory, a branch of mathematics that studies graphs-structures made up of points called vertices connected by lines known as edges. Graph theory has broad applications, including modeling networks such as the Internet. The conjecture, originally proposed in the 1970s by various mathematicians, posits that nearly every graph can be covered by a special set of loops called cycle double covers. A cycle is a closed loop within a graph, returning to its starting vertex without retracing any edge. A cycle double cover is a collection of such cycles that together cover every edge in the graph exactly twice.
Despite decades of effort by prominent mathematicians, the conjecture had only been proven for specific types of graphs, never in full generality. The AI-generated proof released last Friday appears to have settled the problem by demonstrating that any graph to which the conjecture applies can indeed be covered doubly by no more than eight well-chosen cycles. The proof excludes graphs that have large sections connected by a single edge-a technical condition that avoids certain problematic cases, such as two cities connected by a single road.
What makes this achievement particularly striking is that the AI's approach did not rely on groundbreaking new mathematical concepts. Instead, it synthesized and extended existing human methods, extracting more from them than had previously been achieved. This pattern-AI rediscovering or refining known techniques to solve "hard" problems-is becoming increasingly common as companies like OpenAI explore mathematical problems as benchmarks for the reasoning capabilities of their models.
Princeton mathematician Noga Alon praised the accomplishment, calling the conjecture "well-known" and highlighting the surprising brevity of the proof. He described this success as yet another example of how AI tools are already transforming mathematical research. Similarly, Andrew Sutherland, a mathematician at MIT, suggested that some problems earn reputations for difficulty that may discourage human researchers from working on them extensively, creating a sort of self-fulfilling prophecy. AI models, unburdened by such perceptions, might find relatively straightforward solutions to problems long deemed intractable.
OpenAI also disclosed the specific prompt that guided GPT-5.6 Sol toward solving the conjecture, providing insight into the engineering strategies behind the breakthrough. The prompt included what Sutherland called "scaffolding" designed to encourage the AI to invest the sustained effort necessary for a rigorous proof. Notably, the model was instructed to delegate tasks among up to 64 parallel agents, a technique that facilitates collaboration and helps reduce errors such as false proofs or hallucinated references. The prompt explicitly forbade the AI from prematurely concluding that the problem was unsolvable due to its historical difficulty, and it urged the model to persist for at least eight hours before considering giving up.
This method of prompting reflects experiences from mathematicians who have worked with large language models, where AIs often attempt to evade producing novel proofs by citing previous human failures. The key appears to be a combination of encouragement and firm direction-much like a schoolteacher motivating a student to keep trying despite setbacks.
The implications of this breakthrough extend beyond a single mathematical conjecture. It signals a shift in how mathematical research might be conducted in the future, with AI tools taking on increasingly sophisticated reasoning tasks and complementing human intuition. As scientific and technological advances continue to accelerate, collaborations between human mathematicians and AI may unlock solutions to problems that have long resisted traditional approaches.
Joseph Howlett, the reporter covering this story for Scientific American, notes that this development comes at a pivotal moment for science journalism and research. The publication encourages readers to support its mission of delivering impactful science stories, emphasizing the importance of sustaining coverage that highlights discoveries shaping our understanding of the universe.
In sum, OpenAI's GPT-5.6 Sol has demonstrated that artificial intelligence can tackle and solve complex mathematical problems that have eluded human proof for decades. By harnessing a strategic prompting approach and leveraging multiple AI agents working in parallel, the model produced a concise and elegant proof of the cycle double cover conjecture. This achievement not only marks a milestone in AI's mathematical capabilities but also foreshadows a future where AI-assisted reasoning becomes an integral part of scientific discovery.