cyphersolver

An informal benchmark of frontier AI models on historically “unsolved” ciphers, as much a hobby as a measurement. Most of the work is by Claude (Opus 5 and 5.5, some Fable 5.1); a few results came from outside contributors using other models. Each target’s profile.json records which model did what, and what it was given. The first targets came from three standard lists: S. Tomokiyo’s Unsolved Historical Ciphers, Klaus Schmeh’s Top 50 unsolved encrypted messages, and Elonka Dunin’s famous unsolved codes. Since mid-September 2026 most come from this project’s own catalogue of undeciphered archival letters (DECODE records, BnF and Gallica notices, Tomokiyo’s “undeciphered” letters, printed calendars). Each one is a test of how far an AI assistant, working with a human, can get on a problem that has resisted people for decades or centuries: how much is archival research, how much is cryptanalysis, and where it stops. Method: archival research, historical cribs, printed editions, and small…