Claude cracks a 25-year NSA number mystery, hailed as “a real discovery”

CoinDesk reports that Anthropic’s Claude has cracked a number puzzle tied to the NSA that had gone unsolved for 25 years. The headline leans on a quote — “Wow, it’s a real discovery” — to signal how the result landed with the people closest to it.

What the report describes is a concrete outcome rather than a demo: a numerical problem that stood for a quarter of a century now has an answer produced by a general-purpose model. The two qualifiers in the headline, 25 years and the NSA, are what make the case notable — the provenance and difficulty of the problem are far from a routine benchmark task.

The tone of the coverage matters as much as the claim. The surprise in the quoted reaction is aimed at the result itself, not at a score on an evaluation set. Assessments that come from inside the problem domain carry more weight than leaderboard movement, because they speak to whether something has actually been settled.

Public detail is still thin, however. The exact form of the puzzle, the reasoning path Claude took, and whether the answer has been independently checked are all questions that need a fuller technical account. A single case persuades only to the extent that it can be reproduced and verified.

Even so, the case carries a signal. Most public examples of large models in mathematical work so far have cast them as assistants to human researchers. If a model can independently push forward a long-dormant problem, its place in the research workflow shifts from efficiency tool to problem solver.

The things to watch are straightforward: a detailed post-mortem, third-party confirmation of the solution, and whether comparable open problems can be advanced in the same way. If those arrive, the “real discovery” line will read less like a curiosity and more like a measured data point on where model capability now ends.