We assess the system card for Anthropic’s Claude Fable 5 and Mythos 5 models, comparing it against four earlier Anthropic cards and the OpenAI GPT-5.5 and Google Gemini 3.1 Pro cards, with every claim tied to verbatim quotation. The card’s most favourable results are hard for an outsider to reproduce: they rely on internal harnesses, an in-house grader, large token budgets, and runs the public API cannot repeat. The same pattern appears in both competitor cards, however, and on competitor comparison and contamination disclosure Anthropic is the more transparent of the three. The findings that carry weight are the shipped model’s disclosed safety regressions, including a roughly sevenfold rise in how often it continues sabotaging safety-research tasks and a deployed prompt-injection defence worse than two predecessors; the GPT-5.5 card shows close parallels, so these are frontier-wide. We separate the few choices specific to this card from the practices common across the field, and set aside several circulating criticisms (on memorization curves, benchmark omissions, and conflicted external testers) that do not survive checking.