Claude Fable 5.1 cracks 370-year-old royalist cipher

The article doesn't explain what cryptanalytic capability Fable 5.1 actually possesses that prior models lacked, or whether this was a domain-specific fine-tune versus general reasoning. The framing as a 'breakthrough' obscures whether the model solved a genuinely hard problem or simply had better training data coverage of 17th-century cipher techniques.

This lands awkwardly against Anthropic's Fable 5.1 positioning from early September. The release was marketed on scientific reasoning (52.6% on Terminal-Bench-Science) and cost efficiency for agentic workflows, not cryptanalysis. The royalist message decode feels like a post-hoc demonstration of capability rather than a planned use case. More relevant is the SCILAWS-BENCH paper from the same week, which directly challenges whether LLMs can solve novel problems or just pattern-match on training data. That framework would be the right lens to evaluate whether this cipher crack represents genuine discovery or memorized decryption techniques.

If Anthropic or The Decoder releases the actual ciphertext and decryption method, check whether the solution relies on brute-force pattern matching against known historical documents in the training set versus genuine cryptanalytic reasoning. If independent researchers can replicate the decode on similar-era ciphers the model has never seen, that confirms capability; if it fails, this was likely a lucky hit on a specific historical artifact.