- Claude Opus 4.6 discovered a construction pattern that solved a directed Hamiltonian cycle decomposition problem Donald Knuth had worked on for over 30 years.
- The solution was found in roughly one hour across 31 guided explorations, then computationally verified for all odd values of m up to 101.
- Knuth wrote the formal mathematical proof himself, calling the result “a dramatic advance in automatic deduction.”
- The discovery involved human-AI collaboration, not autonomous problem-solving — Filip Stappers guided Claude through the exploration.
What Happened
Donald Knuth, the 87-year-old computer scientist widely regarded as one of the founders of modern algorithm analysis, has seen one of his long-standing open problems solved with the help of an AI model. The problem — a directed Hamiltonian cycle decomposition conjecture — had resisted solution for more than three decades.
Filip Stappers, a colleague of Knuth, posed the problem to Anthropic’s Claude Opus 4.6 hybrid reasoning model and guided it through a series of 31 explorations. Within roughly one hour, Claude identified a “serpentine” construction pattern that cracked the conjecture. Knuth then wrote the formal proof himself.
“It seems I’ll have to revise my opinions about ‘generative AI,'” Knuth said after reviewing the result. He added: “What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction.”
Why It Matters
This is one of the clearest examples of AI contributing to genuine mathematical discovery rather than simply verifying known results or generating plausible-sounding proofs. The problem was not a toy benchmark but an open conjecture from one of the most respected figures in computer science. Knuth authored “The Art of Computer Programming,” a multi-volume work that has shaped the field for over five decades.
The result also highlights a collaborative model for AI-assisted research. Claude did not autonomously solve the problem. Stappers guided the exploration across 31 separate sessions, posing sub-questions and refining the search direction, while Knuth wrote the formal proof. The AI’s contribution was identifying the construction pattern — a task that required exploring a large combinatorial space quickly and recognizing structural relationships that would take a human researcher far longer to enumerate.
Technical Details
The conjecture involved decomposing all edges of a directed graph with m-cubed vertices into exactly three Hamiltonian cycles, for all odd values of m greater than 2. Each vertex in the three-dimensional grid, labeled (i, j, k), has exactly three outgoing arcs that increment i, j, or k modulo m.
Claude discovered a serpentine pattern based on s = (i + j + k) mod m, which independently corresponded to the classical modular m-ary Gray code. The model found 4,554 valid decompositions in total, with 760 involving generalizable cycles that extended to arbitrary odd m.
The solution was computationally verified for all odd m up to 101, providing strong empirical evidence before the formal proof was completed. Claude Opus 4.6, released in February 2026, is Anthropic’s hybrid reasoning model capable of extended exploration across complex problem spaces. The model’s ability to maintain coherent mathematical reasoning across 31 sequential explorations — each building on prior results — was central to the discovery.
Who’s Affected
Mathematicians and computer scientists working on combinatorial graph theory now have a resolution to a problem that had been open since the early 1990s. The result also has implications for researchers studying Hamiltonian decomposition in directed graphs, a topic with applications in network design and error-correcting codes.
AI research labs will likely point to this as evidence that large language models can contribute meaningfully to mathematical reasoning, not just pattern matching or text generation. Google DeepMind’s AlphaProof and AlphaGeometry have tackled competition math, but those systems were purpose-built. Claude is a general-purpose language model, making this result notable for its breadth of capability.
However, the human-guided nature of the discovery tempers claims about fully autonomous AI mathematicians. Stappers’s expertise in knowing which sub-problems to pose was essential to the outcome.
What’s Next
Knuth’s formal proof is expected to appear in a forthcoming publication. The post announcing the result received 635,000 views and 6,000 likes within hours of its March 17, 2026 publication, signaling broad interest from both the math and AI communities.
The key limitation remains that this was a guided collaboration, not an autonomous discovery. Stappers knew which questions to ask, and Knuth provided the mathematical framework. Whether AI models can identify open mathematical problems on their own, formulate the right sub-questions, and produce rigorous proofs without expert human direction remains an open question. For now, the Knuth result demonstrates what human-AI collaboration can achieve when both sides bring complementary strengths.
