Claude Fable 5.1 Solved a 370-Year-Old Cipher — Here's How

A researcher at Vals AI gave Claude Fable 5.1 an open-ended goal — find an unsolved historical cipher and solve it — and in 44 minutes and roughly 176,000 tokens, with zero human intervention, it cracked a 370-year-old cryptogram that had resisted frequency analysis, substitution attempts, and homophonic-cipher approaches since it was first flagged as an open problem in 1899.

The cipher is the Cyphral Distich, a 64-number cryptogram printed at the end of Sir Thomas Urquhart's 1653 book Logopandecteision. It's listed among cryptography researcher Klaus Schmeh's Top 50 unsolved encrypted messages, and the writeup, published by Vals AI researcher Geby Jaff, went viral on Hacker News with over 260 points and 80+ comments.

TL;DR

The clue everyone else missed

Urquhart's cryptogram sits immediately after a section of his book called the "32 Proquiritations" — and Urquhart himself, per the researcher's writeup, "goes out of his way" to emphasize the number 32 repeatedly, alongside language about "wishes" and the reader finding "his own heart's wishes, and the Author's minde." Put together: 32 Proquiritations, 32 numbers in each of the cipher's two lines, and repeated thematic emphasis on wishing.

Most historical attempts, per the researcher, assumed the key had to be external — some reconstructed cipher alphabet or number-to-word mapping existing outside the text. Fable 5.1's breakthrough was realizing the key was internal: the book itself is the cipher key. For the i-th number in a cipher line, go to the i-th Proquiritation, use that number as a word index within it, and take that word's first letter. Applied across all 64 numbers, the result reads: "O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND."

The solution is strongly self-verifying: both lines land at exactly 32 letters (matching the cipher's own structure), the text forms a rhyming couplet ending in "and/land" consistent with the promised "distich" (a two-line verse form), and the content — a prayer for the restoration of Charles II — matches Urquhart's well-documented Royalist politics.

A second, harder puzzle solved the same way

Urquhart left a larger, similarly unsolved cryptogram — the Cyphral Octastich — in his 1652 book The Jewel, consisting of 285 numbers instead of 64. Applying the same underlying logic (numbers indexing into a specific page rather than a specific paragraph, since The Jewel has exactly 284 numbered pages matching the 285 cipher numbers), Fable 5.1 decoded an eight-line royalist prayer in ottava rima (an ABABABCC rhyme scheme), closing with "AMEN, SO BE IT."

The researcher's own writeup is careful about the caveats here, which is worth taking seriously rather than glossing over: 231 of 275 readable positions are exact first-occurrence hits against the source text; the other 44 are off by 1–3 words for identifiable reasons (hyphenated words at page breaks, an untranscribed Greek phrase, and similar transcription artifacts). Nine letters in one line remain genuinely unresolved — the researcher tried page-shift, dropped-letter, and dictionary-lattice hypotheses and found no version that produces readable English, concluding either Urquhart made an error there or the source numbers were misprinted. There's also an unresolved page-shift of one starting partway through the poem, and no freely available scan exists of the physical final quire needed to fully confirm the remaining ambiguities — the researcher flags that a physical 1652 copy or the 1983 Jack & Lyall scholarly edition would be needed to close the gap completely.

Why frequency analysis and substitution never worked here

It's worth explaining why more than a century of conventional cryptanalysis missed this, since the "obvious" answer in hindsight is easy to state but the reason it stayed hidden is genuinely instructive. Standard historical cipher-breaking techniques — frequency analysis, simple and homophonic substitution — all assume the cryptogram encodes a message using a fixed, internally consistent mapping from symbols to letters, independent of any external document. Under that assumption, a 64-number sequence with values ranging from 1 to 70 looks like it should map to a 64-70 character alphabet or a word list, and researchers spent over a century trying to reconstruct exactly that kind of external key.

The actual scheme is a book cipher variant, where the "key" isn't a static alphabet at all — it's a pointer into a specific, already-published external text, with the sender assuming the recipient has access to that same text. Frequency analysis is structurally blind to this category of cipher because the visible numbers aren't standing in for letters directly; they're standing in for positions inside another document, and no amount of statistical analysis of the numbers themselves reveals that without independently guessing which document and which indexing scheme. Urquhart's own repeated emphasis on the number 32 and the word "wishes" functioned as an authorial hint pointing at the correct external document — a clue previous human researchers apparently read past, likely because 19th- and 20th-century cryptography scholarship approached the puzzle with cipher-analysis tools rather than close literary reading of the surrounding text.

How the model was pointed at this problem

The elicitation method is arguably as interesting as the solve itself. The researcher's approach, in their own words: give the model an open goal (solve an unsolved cipher), some encouragement, a pointer to the model's own past math-solving feats as a confidence anchor ("something like this should be easy in comparison"), and instructions to "think creatively and really analyze the problems it encountered." Two explicit constraints shaped the search: avoid ciphers whose creators are long dead and whose puzzles may have no single verifiable answer, and avoid ciphers at the absolute difficulty ceiling — Kryptos K4 was named explicitly as too hard for current models given how much prior expert and institutional effort (including from the CIA) has already failed against it.

Fable 5.1 then searched among candidate unsolved ciphers on its own and identified the Cyphral Distich as unusually tractable — a judgment call the researcher credits as much of the achievement as the actual decoding. As the researcher puts it: "The answer was simple in hindsight. It just kept looking until it found it." That's a notably different kind of capability claim than a benchmark score — it's closer to what explainx.ai covered in Anthropic's Claude proving Fermat's Last Theorem's special-case Lean formalization and Fable 5's Jacobian conjecture counterexample — persistence and search over a large possibility space, rather than a single flash of "cleverness."

The skepticism, and why it matters

The strongest pushback isn't about the model's capability — it's about verification. Several commenters on Hacker News pointed out that no freely available page image of the relevant 1652 edition exists, meaning the ~9 unresolved letters and the page-shift ambiguity can't currently be independently confirmed without a physical copy or the scholarly 1983 edition. Others questioned why a genuinely notable 370-year-unsolved mystery generated so little visible prior academic discussion or news coverage before this writeup — a fair question, though it's also consistent with the researcher's broader point that this specific category of historical puzzle was bottlenecked by human attention, not fundamental difficulty: nobody with the time and inclination to spend days tracing 17th-century printing conventions had gotten around to trying this particular approach.

That's the actual generalizable claim worth taking from this story, separate from any one cipher: a large class of historically unsolved problems — obscure archival puzzles, forgotten mathematical conjectures, and similar tasks — were bottlenecked less by raw intractability than by the scarcity of humans willing to spend days reading obscure material and testing unpromising leads. If that bottleneck is genuinely loosening, it's a meaningful capability shift independent of whether this specific 370-year-old cipher turns out to need one more round of physical-archive verification.

Related reading

• Anthropic's Claude Formalizes a Fermat's Last Theorem Special Case in Lean
• Fable 5 Finds a Counterexample to the Jacobian Conjecture
• Claude Fable 5.1 / Mythos 5.1 Launch: Benchmarks and Pricing
• Near AI's PutnamBench Result: $111, 250x Cheaper
• Zhi Wei Sun's Prime Gaps World Record With GPT-5.6-Sol
• What Is Recursive Self-Improvement (RSI) in AI?

Official source: Vals AI research writeup and the Cyphral Distich page listing at HCPortal

This post reflects the Vals AI researcher's writeup and public discussion as of September 14, 2026. Physical-archive verification of the Cyphral Octastich's remaining ambiguous positions was still pending at time of publication.