Anthropic's Claude Fable 5.1 has solved a numerical cipher that had gone uncracked since 1653, according to the AI evaluation firm Vals AI. The puzzle, known as the "Cyphral Distich," was published by Sir Thomas Urquhart in the 17th century and consists of two lines of 32 numbers each. Vals AI says the model produced a verifiable solution in 44 minutes, with no human assistance.
The result, first reported by The Decoder, is a small but striking data point in an ongoing debate: can frontier models do more than answer benchmark questions — can they actually push forward scholarship that humans have failed to finish? For readers tracking the latest developments in AI research, the answer increasingly appears to be yes, provided the model is pointed at the right problem.
A message hidden for 373 years
The Cyphral Distich has taunted puzzle-solvers for centuries. The mechanism, once revealed, is elegantly simple: each of the 32 numbers in the two lines points to a word in one of the 32 sections of Urquhart's publication. Take the first letter of each selected word and a hidden declaration of political loyalty emerges:
"O God uphold King Charls the Second and make him the supreme ruler of this land."
The message sat in plain sight inside Urquhart's own book. The barrier was never storage or secrecy — it was the combinatorial guesswork of working out which word each number indicated, with no confirmation available until the entire sequence fell into place.
Months of failure, then a 44-minute solve
The breakthrough did not come from a straightforward challenge. According to Vals AI, the team had spent months testing frontier models on unsolved puzzles, and none of the other models produced a verifiable solution to any of them. Fable 5.1 was given a different, more open-ended brief: find a solvable puzzle on its own.
The model sifted through candidate problems, flagged the Cyphral Distich as promising, and then went to work. Forty-four minutes later it had a complete, checkable solution — the first verifiable crack of the cipher in its 373-year history.
Persistence, not genius
The most sobering detail in the report is how the win happened. Citing Vals AI, The Decoder notes that Fable 5.1 succeeded not through superior cryptanalysis or a clever mathematical insight, but through systematic trial and error and sheer persistence. In hindsight, the solution was simple enough that a determined human could have found it too.
That framing matters. The model's advantage was stamina — the ability to grind through hypothesis after hypothesis without fatigue, frustration, or the creeping conviction that a puzzle is unsolvable. Human codebreakers abandoned the Cyphral Distich not because they lacked the intelligence to solve it, but because centuries of failed attempts offer no reward for one more try. An AI agent feels no such discouragement.
What it says about AI as a research tool
Historians, classicists, and archivists have increasingly experimented with machine learning to read damaged manuscripts, reconstruct fragmented texts, and test interpretations at scale. The Cyphral Distich episode adds a new twist: the AI was not handed a pre-formatted data problem. It was asked to survey open problems, judge which one might yield, and then commit to it — a workflow that looks much more like independent research than assisted calculation.
The economics of that workflow deserve attention. For a human scholar, a decade of dead-end attempts on a cipher is a career spent in the wilderness. For an AI agent, failed hypotheses cost electricity and tokens — and a single success can retroactively justify every dead end. That asymmetry changes which problems are worth attacking at all. Puzzles like the Cyphral Distich, once dismissed as objectively unsolvable because no human could sustain the search, suddenly fall into the "plausible weekend project" category. Libraries and archives are full of material with exactly this profile: tightly bounded, verifiable, and abandoned for lack of stamina rather than lack of importance.
A distich, for those who skipped Latin poetry class, is simply a two-line unit of verse — Urquhart's contribution was to replace the poetry with two lines of pure number, daring posterity to reconstruct the text they encoded. The Decoder's report does not say why Urquhart concealed the message, though a groveling prayer for the king's supremacy leaves little to the imagination about which side he backed in a dangerous century.
There is a commercial subtext as well. The result lands days after Anthropic's launch of Fable 5.1, a model the company positions for coding and research work at a significantly lower price point for agentic use. A publicized demonstration of autonomous puzzle-solving is, among other things, a well-timed proof of that marketing pitch.
The verification question also remains open in one respect: Vals AI published the solution and the reasoning behind it, which is what separates a genuine crack from a plausible guess. The words spelled out by the cipher form a coherent, grammatical sentence exactly matching the structure the numbers implied — the kind of constraint that accidental solutions do not survive. In an era when AI outputs are routinely greeted with skepticism, the fact that this one is checkable letter by letter is what makes it worth reporting.
None of this means language models are about to replace domain experts. The cipher's solution still had to be verified by humans, and the model's method was brute-force patience rather than understanding. But as a demonstration of where AI-assisted scholarship is heading — models that choose their own targets and work unsupervised until something cracks — it is hard to imagine a more literary emblem than a royalist riddle from 1653 finally giving up its secret to a machine.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
