Claude cracks 370-year-old cipher

A cipher story with a verification problem

The working headline is simple: Claude cracks 370-year-old cipher. The current state is more complicated, and more interesting. Over the past three days, the Cyphral Distich has moved from a neat AI triumph to a case study in how quickly frontier models can generate plausible discoveries, and how slowly the surrounding evidence can be checked.

The claim centers on Claude Fable 5.1, an Anthropic model used by Vals AI in an autonomous research-style run. According to recent coverage, the model was given an open task: solve Sir Thomas Urquhart’s Cyphral Distich, a two-line numerical cryptogram associated with Urquhart’s 1653 Logopandecteision [1]. The reported performance is the attention-grabbing part: about 44 minutes, 176,000 tokens, and no human interjections during the run [1]. A later roundup presented the same core numbers and framed the result as Claude reading the book rather than brute-forcing the cipher [3].

The proposed plaintext is a royalist prayer for Charles II: “O God uphold King Charles the Second and make him the supreme ruler of this land” [3]. In narrative terms, it fits. Urquhart was a royalist; a hidden devotional-political couplet would not be wildly out of character. In technical terms, the alleged method is also elegant. The numbers are said to point back into the surrounding text: for each position, take the corresponding Proquiritation, count to the indexed word, and use that word’s first letter [1].

That is why the story traveled. It has everything an AI breakthrough anecdote needs: an old unsolved puzzle, a frontier model, a short runtime, a humanly legible answer, and a “why didn’t we see it?” rule. But the freshest reporting also makes clear that the decryption is not yet the same thing as a settled historical result.

What Claude appears to have found

The key idea is that the Cyphral Distich may not be a substitution cipher at all. Instead of mapping numbers to letters through a detached alphabet, the numbers allegedly use Urquhart’s own text as the key [1]. This matters because many classical attacks, including frequency analysis, are weak against a very short message. With only 64 numbers, there is not much statistical surface to grab.

A book cipher changes the problem. The solver must discover not only what the numbers mean, but which text they index, how the text is divided, how words are counted, and whether spelling, punctuation, Latin phrases, hyphenation, or transcription differences affect the count. For an AI model, that is a natural search space: read, hypothesize, test, reject, and repeat. That is the part of the claim that feels credible even to skeptics. Modern language models are good at pattern hunting across messy text, especially when the target answer should look like grammatical language.

Recent summaries say Fable 5.1 noticed that the cipher’s two lines each contain 32 numbers and that Urquhart’s relevant surrounding material contains 32 Proquiritations [3]. From there, the rule becomes compact: the first cipher number indexes into the first Proquiritation, the second number into the second, and so on [3]. The model’s alleged answer then produces a 32-letter line followed by another 32-letter line, ending in a rhyme on “and” and “land” [1].

If correct, the achievement is not that Claude invented a new cryptographic attack. It is that it selected a tractable archival puzzle, connected local textual clues, and tested a book-cipher hypothesis fast enough to produce a coherent answer in less than an hour. That is a real capability signal, but not the same signal as “AI has defeated cryptography.” The cipher was difficult because of attention, provenance, and indexing, not because it resembled a modern cryptosystem.

The counterclaim: beautiful answer, missing receipts

The story’s current status changed because independent checking quickly entered the picture. Forbes reported on September 15 that Reticuli Labs had published a replication attempt saying the published method could not reproduce the claimed plaintext from the surviving 1653 text it examined [1]. DrafterDaily’s September 16 analysis went further, emphasizing that the claim, the refutation, and the viral spread occurred on different timelines: Vals AI first published the proposed solution on August 31, Reticuli answered on September 1, and the claim then surged through Hacker News-style circulation around September 14 [2].

The core criticism is not vague skepticism. The refutation, as summarized by recent coverage, says the relevant 1653 book copy does not contain the numeric distich where the claim places it, and that ten of the 64 required characters cannot be generated from the corresponding Proquiritations under the described rule [2]. DrafterDaily also reports that Reticuli ran a grid of 65 counting conventions and found a best result of 8 matches out of 64, roughly chance-level for this kind of first-letter test [2].

That creates a sharp fork. Either the Vals run used a different source text, edition, transcription, or page tradition that really does contain the needed words in the needed positions, or the model generated a highly plausible but non-reproducible solution. The difference is not philosophical. It can be settled by publishing the exact scans, transcriptions, scripts, and counting conventions used.

A separate September 16 check complicated the provenance question in another way. Paper.wf’s Theodore note argued that the Cyphral Distich is not a hallucinated modern artifact, but that people looking for it in the 1653 scan may be looking in the wrong place; it points to an 1899 Notes and Queries printing and to John Willcock’s 1899 biography as key waypoints in the record [5]. That does not validate Claude’s plaintext. It does show why the bibliographic trail matters: if the cipher’s printed history runs through later reproductions, then “the original book” may not be enough to verify the exact numerical text or its intended key.

Why the case matters for AI reasoning

The Cyphral Distich is useful precisely because it sits between benchmark and archive. It is small enough for outsiders to understand and, in principle, check. It is also messy enough to expose the difference between producing an answer and establishing one.

For AI labs, the positive interpretation is obvious. Claude Fable 5.1 appears to have behaved like a tireless research assistant: searching candidate problems, discarding unproductive approaches, recognizing a structural clue, and forming a compact hypothesis. Recent Vietnamese and Korean reposts of the original claim repeated the same headline ingredients: 64 numbers, 32 text units, 44 minutes, 176,000 tokens, and a book-based first-letter rule [4][5]. The global spread of those summaries shows why the story resonated beyond cryptography circles.

For researchers, the caution is just as important. AI can now generate candidate discoveries cheaply. Verification remains expensive. A model can produce a rhyming royalist couplet before lunch; a human checker may still need to inspect scans, editions, transcriptions, and historical references for days. Forbes framed this as a broader shift: AI may flood researchers with cheap hypotheses while leaving humans with the hard job of deciding which ones are real [1].

That is not a failure mode unique to Claude. It is a structural issue for every agentic research system. When the output is code, tests can run. When the output is a proof, specialists can inspect the proof. When the output is a claim about a 17th-century printed artifact, verification depends on the artifact, not the fluency of the answer.

The honest verdict today

As of September 17, 2026, the safest formulation is this: Claude Fable 5.1 has produced a vivid, specific and potentially important proposed solution to the Cyphral Distich, but the result remains contested in the freshest reporting [1][2]. The original claim says the model solved the cipher in 44 minutes; the strongest recent skeptical account says the published method does not reproduce the plaintext on the examined source material [1][2]. The next decisive step is not another opinion piece. It is the release, or independent inspection, of the exact source pages and counting procedure.

If the solution holds, the story will be memorable: a 17th-century cryptographer’s hidden royalist couplet finally surfaced by a language model with enough patience to test the right hunch. If it fails, the story will still be memorable, but for a different reason: it will show how close a modern model can get to a historically plausible “discovery” before the archive says no.

Either way, the Cyphral Distich has already become a better test than the headline promised. It is not only about whether Claude cracked a cipher. It is about whether the AI era can learn to publish the receipts as quickly as it publishes the miracle.

Sources from the last 72 hours

1. [1]Did AI Crack A 370 Year Old Cipher?Sep 15, 2026, 1:31 PM UTC
2. [2]The Cipher Was Solved on 31 August. It Was Refuted on 1 September. It Went Viral on 14 September.Sep 16, 2026, 12:00 AM UTC
3. [3]Claude Fable Solves A 370-Year-Old Cypher In 44 MinutesSep 15, 2026, 12:00 AM UTC
4. [4]Fable 5.1 giải mật mã Cyphral Distich tồn tại 370 nămSep 14, 2026, 8:07 PM UTC
5. [5]클로드 페이블 5.1이 키프랄 디스티치를 풀었다Sep 16, 2026, 1:33 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.