Claude Fable 5.1 reportedly cracks Sir Thomas Urquhart's 370-year-old Cyphral Distich

A centuries-old cipher falls to a single model run

According to Vals AI, the AI model Claude Fable 5.1 has solved the Cyphral Distich, a cryptogram by the 17th-century Scottish author Sir Thomas Urquhart that had resisted decipherment for roughly 370 years. The company says the model needed 44 minutes, about 176,000 tokens of reasoning, and no mid-run human input. The blog post, published on September 13 and later promoted to Hacker News's front page, further claims the model decoded all but nine letters of a second, larger Urquhart cryptogram.

The puzzle

The Cyphral Distich appears at the end of Urquhart's Logopandecteision: two lines of 32 numbers each, 64 figures in all, and nothing else. Vals AI traces its history as an open problem, noting it was posed in Notes and Queries in 1899, revisited in 20th-century cryptography literature, and listed by cipher researcher Klaus Schmeh among his top 50 unsolved encrypted messages. Earlier attempts leaned on standard techniques such as frequency analysis, substitution and homophonic substitution. All failed, the post argues, because they assumed the key was an external mapping, some cipher alphabet or number-to-letter table that had to be reconstructed from outside the text.

The insight the humans missed

Fable 5.1's solution rested on two observations about the book itself. First, the cryptogram sits directly after Urquhart's 32 Proquiritations, a count Urquhart pointedly underlines in the text ("there can no number like that of two and thirty … be pitched upon"). Second, the poem accompanying the cipher promises the honest reader will find "his own heart's wishes, and the Author's minde," and the Proquiritations repeatedly end with language about wishes and desires.

Put together, 32 Proquiritations, 32 numbers per cipher line, and repeated talk of wishes, the rule falls out: for the i-th number in a line, go to the i-th Proquiritation, count into it by that number, and take the first letter of the word you land on. The recovered plaintext reads:

O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND

Vals AI calls the result strongly self-verifying: each line is exactly 32 letters, the two lines rhyme as a distich should, and a covert prayer for Charles II fits what is known of Urquhart's Royalist politics.

A second cryptogram, mostly solved

The same run reportedly tackled the Cyphral Octastich in Urquhart's The Jewel (1652): 285 numbers across eight lines and a closing "decagram." Here the key is pagination. The Jewel has 284 numbered pages, and the k-th number indexes a word on page k, again taking its first letter. The decoded text is an ottava rima royalist prayer that Vals AI dates to London in March 1652, asking God to sustain "that Regal Familie whereof King Charls the Second is the Head" and closing with "AMEN, SO BE IT."

Verification looks solid on most positions: 231 of 275 readable spots match the first word on the relevant page in the EEBO-TCP transcription, with the remaining 44 off by one to three words for identifiable reasons such as hyphenated words at page tops.

Caveats

Vals AI is explicit about loose ends. Nine letters, chiefly a run on line 5 that decodes to "CONERTHTO," resist interpretation; the author tested page-shift, dropped-letter, misprint and dictionary hypotheses without producing English. The word IRISH appears to be enciphered without its second vowel, either a slip or a metrical contraction. From position 159 onward every number keys to the previous page, suggesting one page was reused or a figure was duplicated in print. And because no freely available page images of the 1652 edition exist online, final confirmation of the near-miss positions and the nine unresolved letters requires a physical copy or the 1983 Jack & Lyall edition. Verification scripts were released alongside the post.

How the model was steered

The author says months of attempts to elicit a verified solve from frontier models had failed before this run. The prompt itself was light: a goal, some encouragement, references to the model's past mathematical results, and two constraints. Avoid ciphers with existing solutions or many plausible readings, and avoid the hardest targets such as Kryptos K4. The author concedes the decisive clue is simple and that other models might not necessarily fail here, crediting Fable with knowing when to abandon unpromising problems and seizing on this one.

Why it matters

Two things make this more than a curiosity. First, the solve did not come from faster brute force but from reading context: noticing an emphasized number and a repeated motif, then connecting both to the cipher's structure. That is exactly the kind of cross-referencing human cryptanalysts attempted for over a century without success, and it suggests AI models can contribute to long-open problems in the humanities, not just math benchmarks. Second, the case illustrates why verification matters. This cipher was a good target partly because its solution checks itself, while many purportedly unsolved ciphers may never be confirmable at all. The claim still rests on a single organisation's post and awaits scholarly confirmation, but with the method and scripts published, checking it is now mechanical rather than mysterious.