A 372-year-old puzzle just fell to a machine. Anthropic’s Claude Fable 5.1 decoded the “Cyphral Distich,” a cryptic two-line sequence of 32 numbers that Sir Thomas Urquhart published back in 1653. The solution, according to Vals AI, came in 44 minutes with zero human assistance.
The message itself is a royalist plea: “O God uphold King Charls the Second and make him the supreme ruler of this land.” The numbers, it turns out, weren’t a mathematical cipher at all. Each one pointed to a specific word in one of the 32 sections of Urquhart’s own book. Take the first letters of those words and the sentence assembles itself. In hindsight, embarrassingly simple.
What makes this interesting isn’t the cryptanalysis. It’s the methodology. Vals AI had spent months throwing frontier models at unsolved puzzles, and none produced anything verifiable. So they flipped the approach: instead of assigning Fable 5.1 a specific puzzle, they let it hunt for one it considered solvable. The model sifted through candidates, flagged the Cyphral Distich as promising, and then ground through trial and error until the answer emerged.
That’s the part worth sitting with. The model didn’t win because it’s a better codebreaker than human cryptographers. It won through brute persistence and systematic elimination. For any organization staring at a stack of intractable problems, the lesson isn’t that AI has some mystical insight — it’s that relentless iteration on the right problem beats cleverness on the wrong one. The bottleneck isn’t the model. It’s knowing which problem to point it at.
