Glonce

Fable 5.1 solves a cipher unsolved for 370 years

agentsThe Glonce editorsHacker News🔥78Glonce rating78Impact50Novelty70Relevance90Surprise75Players35Credibility80

A blog post from vals.ai reports that an AI model it calls Fable 5.1 was set an open task: crack the Cyphral Distich, a cryptogram at the end of Sir Thomas Urquhart's 1653 book Logopandecteision. The cipher is two lines of 32 numbers each, 64 numbers in total. It had gone unsolved since at least 1899, when it was posed as an open problem in the journal Notes and Queries; it resurfaced in 20th-century cryptography writing and was later placed on historical-cipher researcher Klaus Schmeh's list of the Top 50 unsolved encrypted messages. Earlier attempts using frequency analysis, substitution and homophonic substitution all failed, according to the post.

The post says Fable 5.1 produced a solution after 44 minutes and 176,000 tokens of work, with zero human interjections. It tried a few approaches before landing on the insight that broke the cipher open: the key was not an external alphabet or mapping, as most past attempts had assumed, but the book itself. The cryptogram sits right after Urquhart's 32 “Proquiritations,” short numbered passages, and Urquhart draws attention to that number directly, writing “there can no number like that of two and thirty … be pitched upon.” A poem printed beside the cipher promises that a careful reader will find “his own heart's wishes, and the Author's minde,” and the Proquiritations themselves often end in phrases like “is the desire,” “wish” or “hope of.” Combining the two clues gives the rule: for the i-th number in a cipher line, go to the i-th Proquiritation, use that number as a word index into it, and take the first letter of that word. Run across both lines, it yields “O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND,” a two-line prayer for Charles II. The decode checks itself: each line comes out to exactly the promised 32 letters, the lines rhyme on “and” and “land,” and the content fits what is known about Urquhart, a committed Royalist.

Urquhart left a second, larger cryptogram in the same style, the Cyphral Octastich in his book The Jewel (1652): 285 numbers shared with a short accompanying cipher, instead of 64, and likewise unsolved. The post says Fable 5.1 deciphered this one too, with pages standing in for paragraphs. The 1652 Jewel has exactly 284 numbered pages, and the octastich plus a short accompanying cipher called the Decagram carry 285 numbers between them, so the k-th number becomes a word index into page k, and the rule again takes that word's first letter. The result is an eight-line, rhymed prayer (ottava rima, ABABABCC) said to have been written in London in March 1652, again naming Charles II, plus the Decagram's own short line, “AMEN, SO BE IT.” One word in it, “sweigh,” only makes sense as an archaic Scots word for “sway, controlling power”; the post cites a period dictionary entry and an attested idiom from around 1600 for it, and notes that it also happens to rhyme correctly with “familie” and “authoritie.” Checked against a digitized 1652 text (EEBO-TCP), 231 of 275 readable decoded positions land exactly on the first occurrence of the needed letter on the page; the post attributes the other 44 to identifiable transcription quirks, such as hyphenated words at the top of a page or an untranscribed Greek phrase.

Not everything resolved. Line 4 decodes to four letters, I-R-S-H, from pages 127 to 130; the post reads this as either a copying slip for IRISH or a deliberate contraction, since IRSH keeps the line at exactly ten syllables. Nine letters in line 5, positions 4 to 12, pages 149 to 157, decode to C-O-N-E-R-T-H-T-O. Eight of those nine are still exact first-occurrence hits on their pages, so the post says this is genuinely what the digitized text yields, but the result does not spell English, and testing page-shift, dropped-letter, misprinted-number and dictionary-lattice explanations produced no fix; the post concludes Urquhart either erred here or the printed numbers themselves are wrong. From position 159 onward, decoding only works if every page reference shifts back by one, which the post can only guess at: a page reused twice, or a duplicated “5.5” in the printed line. No free page images of the actual 1652 Jewel are available online (the post says a subagent checked Google Books, HathiTrust, the Internet Archive and the National Library of Scotland), so confirming the shift and the nine missing letters would need a physical copy or the 1983 Jack and Lyall scholarly edition.

The post is written in the first person by Geby Jaff, who says they had spent the past few months trying to get AI models to solve an important, still-unsolved cipher, and that no other frontier model tried in that time had produced a verified solve. The method with Fable 5.1 was plain: set it the open goal of solving an unsolved cipher, point it to its own strong past results on math problems as a difficulty anchor, and tell it to reason creatively and really analyze the problems it encountered. Two constraints shaped which problems it could pick: skip ciphers that already have an accepted solution or could fit many equally plausible answers, since a puzzle's long-dead creator cannot confirm a guess; and skip the very hardest cases that thousands of people, or organizations such as the CIA, have already attacked, such as Kryptos K4, which the author judges still too hard for current models. Fable 5.1 reportedly looked over several candidate problems, recognized quickly when one was not going anywhere, and picked up on the Cyphral Distich's clue almost immediately. The author's stated takeaway is that models can now do this kind of work, not just mathematics, because the real bottleneck was always human attention, someone willing to spend hours tracing references and testing ideas that mostly fail; that bottleneck, the author argues, is fading. The author is explicit that neither solve took an extraordinary feat of cryptanalysis. The answer was simple once found.

The post calls the model “Fable 5.1” throughout except for three uses of “Claude Fable 5.1” (the headline, the subtitle and the opening line) and one bare “Claude” near the end; it never says which company or lab built it. It also cites no independent cryptographer or Urquhart scholar who has reviewed either decoded solution.

Key facts

  • Fable 5.1 reportedly solved Sir Thomas Urquhart's Cyphral Distich, a two-line, 64-number cryptogram at the end of his 1653 book Logopandecteision, unsolved since it was posed as an open problem in Notes and Queries in 1899 and later listed among Klaus Schmeh's Top 50 unsolved encrypted messages.
  • The reported key was internal to the book: the i-th number in a cipher line indexes a word in the i-th of Urquhart's 32 Proquiritations, and that word's first letter is the plaintext letter, yielding a 32-letter-per-line prayer for Charles II, “O GOD UPHOLD KING CHARLS THE SECOND AND / MAKE HIM THE SUPREME RULER OF THIS LAND.”
  • Using the same method with pages instead of paragraphs, Fable 5.1 also reportedly deciphered most of a second, larger cryptogram, the Cyphral Octastich and its short accompanying cipher, 285 numbers between them, in Urquhart's The Jewel (1652); nine letters in one line still do not resolve to English.
  • The post says the Distich solve took 44 minutes and 176,000 tokens with zero human interjections, after months in which the author tried other frontier models without a verified solve on any unsolved cipher.
  • The source never identifies who built the model, calling it “Claude Fable 5.1” three times up front (the headline, the subtitle and the opening line) before settling on “Fable 5.1,” and names no independent cryptographer or Urquhart scholar who has checked either decoded solution.

Why it matters

An AI model reportedly cracked, in under an hour and with no one steering it, a cipher that had beaten codebreakers and cipher historians for well over a century. The post is explicit that this was not a display of raw cryptanalytic power: the winning move looked simple once found, nothing like the scale of effort that thousands of people, or organizations such as the CIA, have already thrown at famous holdouts like Kryptos K4, a puzzle the author calls still too hard for current models. The insight, that the cipher's key was internal to the book rather than an external alphabet, had apparently escaped every earlier attempt despite the clue sitting in the text itself. The same approach then extended to a second, larger Urquhart cryptogram, the Cyphral Octastich, which the post says came back almost entirely readable. The author's broader claim is that this kind of long, unsupervised, exploratory work, once bottlenecked by how much attention a human historian or hobbyist could spare, is now something a model can attempt on its own.

Who it affects

The most direct audience is the small community that tracks historical ciphers, including the researchers behind lists like Klaus Schmeh's unsolved-message catalog, plus scholars of Sir Thomas Urquhart and 17th-century Scottish Royalist writing, who now have two candidate plaintexts to weigh. Beyond that niche, the post targets people assessing what current AI models can do on open-ended, low-supervision research work rather than fixed benchmarks: the vals.ai post frames the result as evidence for a new kind of capability, archival and historical problem-solving, alongside the math results it says Fable 5.1 was already known for. Anyone trying to track the model itself is left with little to go on, since the source never names a company or lab behind it.

How to use it

There is no product to buy here, but the post describes a method for pointing a model at this kind of open problem. The author set Fable 5.1 a broad goal, solve an unsolved cipher, cited its own past math results as a difficulty anchor, and told it to reason creatively and really analyze the problems it encountered. Two guardrails shaped which puzzles it could attempt: skip ciphers that already have an accepted solution or could fit many equally plausible answers, since a long-dead author cannot confirm a guess; and skip the very hardest cases already attempted by humans, such as Kryptos K4, judged still out of reach for current models. Within those limits, the post says the model chose among candidate problems itself and recognized quickly when a lead was not paying off. The underlying work is named, not published: a small set of scripts, among them verify_octastick.py and jewel_pages.py, plus solution and status files that the post lists without a link, leaving nothing a reader can actually check.

How solid is it

The two solutions rest on different footing. For the Distich, the evidence is tight: the decoded lines run to exactly the promised 32 letters each, they rhyme on “and” and “land,” consistent with the promised distich, and the resulting prayer for Charles II fits Urquhart's documented politics. For the longer Octastich, the fit is looser. Checked against a digitized 1652 text, 231 of 275 readable positions land on the exact first occurrence of the needed letter, and the post attributes the other 44 to identifiable transcription quirks. But nine letters in one line do not spell English under any of four hypotheses the author tried, even though eight of those nine are still exact first-occurrence hits, and everything past position 159 only decodes after shifting every page reference back by one, a shift the post can only guess at. No free digital images of the actual 1652 printing exist to check further; a physical copy or a 1983 scholarly edition would be needed to close the gap. On top of the cipher-specific uncertainty sits a sourcing one: this is a single, first-person blog post signed by Geby Jaff, with no independent cryptographer or Urquhart scholar cited as having reviewed either claim, and no explanation of who built the model or how “Fable 5.1” relates to “Claude,” a name the text uses for it twice without comment.

Risks and caveats

Nine letters of the Octastich remain unresolved, and the post's explanation for why the sequence shifts by one page partway through is a guess, not a confirmed cause. Line 4's four-letter reading is treated as either a copying slip or a deliberate metrical contraction, and the post picks one without ruling out the other. A closing footnote adds a further wrinkle: one coordinate only produces the reported answer if the Latin phrase “hinc inde” is counted as a single word, otherwise that position shifts by one. More broadly, the two ciphers were not a random test: the author explicitly screened out both already-solved or unverifiable puzzles and the hardest ones humans have tried, including Kryptos K4, which limits how far this particular success says anything about harder cases. And the whole account, including the months of unsuccessful attempts on other models the author says came before it, rests on one source, signed by Geby Jaff, with no outside check.

“The key was not an external cipher alphabet at all. The key was the book itself.”

— the vals.ai post announcing the result