Claude Fable 5.1 Solves 370-Year-Old Cipher in 44 Minutes
An AI just cracked a 17th-century cryptogram that baffled humans for centuries, proving models can now solve obscure archival mysteries.
Models · Source: Hacker News
What happened
Claude Fable 5.1 just cracked the Cyphral Distich. This is a 370-year-old unsolved cryptogram published by Sir Thomas Urquhart in 1653. It took the model exactly 44 minutes and 176,000 tokens to find the answer. The human operator gave the AI an open-ended goal to solve an uncracked cipher. The model was instructed to avoid problems that were too hard, like Kryptos K4, or problems that could not be easily verified. Fable 5.1 evaluated multiple options before selecting this specific puzzle.
Human cryptographers failed for centuries because they assumed the key was an external alphabet. They tried frequency analysis and substitution without success. Fable 5.1 succeeded by spotting an obvious contextual clue. The cipher consisted of two lines of 32 numbers. It was printed immediately after 32 numbered paragraphs called Proquiritations. The model realized the book itself was the cipher key. It used each number as a word index for the corresponding paragraph and extracted the first letter of that word.
The resulting plaintext revealed a hidden Royalist prayer for King Charles II. The solution perfectly matched the two-line rhyming structure promised by the author. Fable 5.1 then applied this exact same logic to a second, much larger cryptogram by Urquhart called the Cyphral Octastich from 1652. It cracked that one too. The model mapped 285 numbers to the 284 pages of the book, proving its initial discovery was not a fluke.
Key facts
- 370 — Years the Cyphral Distich remained unsolved
- 44 — Minutes it took Claude Fable 5.1 to solve the cipher
- 176k — Tokens used by the model during the solving process
- 32 — Numbers in the cipher matching the preceding paragraphs
- 285 — Numbers in the second cipher solved by the model
Why it matters
This changes how we view the limits of AI reasoning and autonomous research. We usually test frontier models on complex mathematics, coding benchmarks, or logic puzzles. This event proves models can now execute open-ended research tasks across highly unstructured historical data. The primary bottleneck for solving obscure problems used to be human attention. Someone had to care enough to spend days reading obscure material and testing unpromising ideas. That bottleneck is now gone. You can point an AI at an archive and let it grind relentlessly.
The second-order effect is a massive unlocking of dead data across all industries. Archives, historical records, and forgotten scientific conjectures are full of patterns humans lack the time or patience to find. If an AI can autonomously select a tractable problem, test hypotheses, and verify its own work without human intervention, every digitized dataset is about to get mined for new discoveries. This shifts AI from a tool that answers questions to an agent that finds the right questions to ask.
For builders
Build autonomous research agents for niche domains
Fable 5.1 succeeded because it was given a broad goal and specific constraints, not a rigid step-by-step guide. Founders can build agents that autonomously hunt for solvable problems in neglected datasets. Universities, legal firms, and research labs will pay heavily for tools that surface overlooked patterns in their archives.
Leverage context windows for unstructured text
The model solved this puzzle by keeping the entire structure of a 17th-century book in its context window. Engineers should stop relying solely on external tools or complex pipelines for pattern matching. Feed raw, unstructured data directly into large context models and ask them to find the hidden rules.
Design self-verifying output loops
The AI chose this cipher because the solution would be self-verifying through rhyme and historical context. Builders must design agent workflows that include built-in verification steps. If the model can independently prove its answer is correct, you eliminate the need for costly human oversight.
My take
Humans overcomplicate things. Cryptographers spent centuries trying complex frequency analysis on this text. Fable 5.1 just looked at the page structure, saw the obvious pattern, and solved it in under an hour. AI does not need to be smarter than us to win. It just needs to be infinitely patient and willing to look at the obvious clues we ignore.
Original reporting: Hacker News. This is my rewrite and opinion.