Claude Fable 5.1 has produced a reading of the Cyphral Distich, a cipher that had sat unsolved inside a publication by Urquhart since 1653. The model was not handed the puzzle. It was asked to go and find an unsolved problem it believed it could crack, worked through candidates, and settled on this one as promising. The key turned out to be the book itself: each number in the cipher points to a word in one of the volume's 32 sections, and the first letters of those words spell a prayer asking God to support King Charles II and make him supreme ruler of the country.
The result arrives after a long run of nothing. The person who set the exercise up had spent several months putting other frontier models against unsolved puzzles, and not one of them returned a solution that could be verified. That is the base rate this sits against: months of failure, then a single hit.
How the hit happened is the part worth reading twice. According to Vals AI, Claude Fable 5.1 got there through systematic trials, working through options and refusing to stop — not through better cryptanalysis. No new technique was brought to the Distich. The model tried, and kept trying, across a space a person would plausibly have given up on.
In hindsight the solution is simple enough that people could have reached it. Looking at a royalist publication from 1653 and asking whether its numbers index words in the same volume is not a leap; it is checking an obvious hypothesis. What nobody did was check it exhaustively.
This reads to me like an endurance result being described in the vocabulary of an intelligence result. The capability on display is the willingness to grind through a large and boring search without losing the thread. That is real, and it is also the capability that scales with money rather than with anything we would recognise as insight. It is the difference between a model that sees the trick and a model that will test every trick until one holds.
The self-selection deserves more scrutiny than it has been given. A model asked to pick its own unsolved problem will pick one it can solve — that is what "promising" means in the instruction it was handed. Claude Fable 5.1 surveyed the field and chose the puzzle whose answer was sitting inside the same binding as the question. That is a legitimate result and a compromised benchmark at once, and the account does not separate the two.
There is also the matter of what counted as confirmation. A search that runs over 32 sections and word positions will throw off a great many letter strings, and something has to decide which of them is a message. For a royalist publication of 1653, a plea for Charles II is close to the most predictable sentence in the language, which makes it both the likeliest true answer and the easiest false positive to fall in love with. Nothing in the account describes what ruled the alternatives out.
That is the tension to carry forward. If what is being demonstrated is persistence rather than technique, the next old cipher to fall will tell us how much compute somebody was willing to spend, not how much better the model got. The puzzles will keep coming apart. What their falling measures will keep shrinking.