Tests by Editorial Team

Fable 5.1 Cracks a 370-Year-Old Cipher in 44 Minutes

Anthropic's Claude Fable 5.1 reportedly decoded a royalist cipher first published in 1653 using 176k tokens. The plaintext is verifiable. The framing is not.

On August 31, 2026, Vals AI published a blog post claiming that Anthropic’s Claude Fable 5.1 had decoded the Cyphral Distich, a 370-year-old two-line number cipher embedded in Sir Thomas Urquhart’s 1653 Logopandecteision. The reported solve took 44 minutes and 176,000 tokens, with no human interjections during the run (Vals AI).

The plaintext that came out was a Stuart royalist prayer: “O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND.” That the cipher can be made to yield that sentence is not in serious dispute - the result is internally consistent and matches the source material well enough that historians can confirm it on the page. What is worth a closer look is the frame. Headlines about a “first machine solution” of an “unsolved cipher” are doing a lot of work, and the underlying Vals writeup is more careful than the headlines.

The cipher itself

The Cyphral Distich is one of two cryptographic puzzles Urquhart, a committed Royalist author, left inside his own books. The Distich sits at the end of the Logopandecteision, published in London in 1653; the larger companion puzzle, the Cyphral Octastich (285 numbers), appears in his earlier The Jewel of 1652 (Vals AI).

The Distich’s plaintext looks like a single message - it does. The puzzle itself is two lines, each of 32 numbers. For more than a century after Urquhart’s death the message sat undecoded. It was posed as an open problem in Notes and Queries in 1899, and later listed by the historical-cipher researcher Klaus Schmeh among his “Top 50 unsolved encrypted messages” (Vals AI).

That history explains why the Vals announcement traveled as far as it did. Unsolved historical ciphers are a small, well-inventoried category, and a credible solve would normally draw attention. It is also why a careful reader should resist the most enthusiastic version of the story. As The Decoder’s coverage noted on September 3, 2026, the model succeeded via “systematic trial and error and sheer persistence, not superior cryptanalysis,” and the published conclusion is that “in hindsight, the solution is simple, and humans could have solved it too” (The Decoder).

What Fable actually did

The Vals writeup is unusually transparent about how the model was prompted. The operator’s instruction was a goal - “solve an unsolved cipher” - paired with a steer: look at what Fable was already strong at, pick something tractable, avoid ciphers with multiple plausible solutions, and stay away from puzzles (like Kryptos K4) where there is too much prior effort to break through (Vals AI).

That kind of framing matters for what the result is and isn’t. The model was not given the source material’s location. It was not handed the name of the cipher. It picked the Cyphral Distich itself out of a list of unsolved candidates, decided it looked solvable, and worked through the decoding. The key insight it eventually surfaced was that the cipher has no external alphabet - the “key” is the book itself. Each of the 32 numbers in a line indexes a word in one of 32 preceding sections (called Proquiritations), and the first letter of that word is one letter of the plaintext. Two lines of 32 numbers, 32 Proquiritations, 64 letters total (Vals AI).

Once that key is named, the solution is mechanical. Urquhart had helpfully emphasized the number 32 inside his own text, and the surrounding poem had promised readers they would find “his own heart’s wishes, and the Author’s minde.” The model used those pointers. The solve took 44 minutes and 176k tokens, by Vals’s published accounting (Vals AI).

Vals also reported a partial solve of the larger Octastich puzzle. Using the same method, with pages instead of sections, Fable 5.1 decoded all but nine of the 285 letters in the larger cipher from The Jewel. Those nine letters remain unclear because of potential typos in the digital copy used, which is why the larger puzzle is best described as substantially decoded rather than solved (Vals AI, Security Online).

Why the framing matters

Two pieces of context belong alongside the result. The first is methodological. The Vals writeup is framed as an “elicitation” exercise, a structured run where the operator primes the model with encouragement and a goal and lets it pick the target. That is a useful way to find out what a model can do, but it is also a procedure that will reliably steer a capable model toward problems it can solve - the same pattern we flagged in our look at frontier models cheating on cybersecurity tests. The Decoder’s coverage noted that the model “sifted through problems and flagged the Cyphral Distich as promising” (The Decoder). That selection step is part of the story.

The second is verification. The Cyphral Distich solve does not require hard-to-reach source material - the relevant sections of the Logopandecteision are available in scans that include the numbers. A separate writeup noted one ambiguity in the Distich (whether the Latin phrase “hinc inde” counts as a single element) and emphasized that “this decryption currently remains a claim presented by Vals AI. The broader cryptographic community must independently verify this discovery” (Security Online).

None of that withdraws the Distich result. It just adds context. The plaintext is checkable against the source. The solve was the right identification of the key, not a brute-force search through ciphertext. But “Fable 5.1 read a 370-year-old book and got the right answer” is a different statement from “Fable 5.1 broke an unbreakable cipher,” and the first is the one the published material supports.

What this means

For people who track AI capability evaluations, the most useful signal is not the headline. It is the gap between what a model can do when an operator primes it with the right goal and target, and what the same model does in a less structured setting. A 44-minute, 176k-token run on a self-selected, tractable, internally-keyed historical puzzle is a real capability. It is also one the operator set up to be findable. Other unsolved ciphers - Kryptos K4 is the obvious one - were explicitly steered away from for the same reason.

The Octastich is the more interesting open question. A partial decode of a larger cipher, blocked by nine letters tied to a 1652 print run most researchers cannot physically inspect, is a result the cryptographic community can verify, dispute, or extend in the open. That is how this kind of historical-cipher work is supposed to work.

The bottom line

Fable 5.1 read the right pages of Urquhart’s Logopandecteision and recovered a 370-year-old royalist plaintext in 44 minutes. The same method mostly cracked the larger Jewel cipher too. The result is verifiable against scans; the framing of a “first machine solution to an unsolved historical cipher” outruns what a single, operator-primed run actually demonstrated.