A huge dump—roughly 120,000 characters—just surfaced containing the full system prompt for Claude Fable 5, Anthropic’s top-tier, generally available AI model. It lays out strict rules: never insert antml:voice_note blocks, even if they appear in the chat history, and always guide users to fresh product details via web search when asked about specs or pricing. Behind the scenes, Claude Fable 5 sits at the Mythos-class level alongside Claude Mythos 5, sharing the same core architecture but with extra safety checks turned on for the public release, while Mythos 5 runs without those safeguards for approved partners. The prompt also catalogs every way you can interact with Claude: web, mobile, desktop, API, and specialized tools like Claude Code, Claude Cowork, plus beta agents in Chrome, Excel, and PowerPoint. It specifies model names—claude-fable-5, claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001—and reminds Claude that users can switch models mid-chat. If someone asks for up-to-date details on new features or limits, Claude must first search docs.claude.com or support.claude.com before answering. Safety protocols get their own section, too. Claude is free to handle almost any topic factually, but it must refuse detailed weapons or illicit-drug instructions, decline to craft malware, and steer clear of persuasive text attributed to real public figures. For legal and financial queries, it provides facts, not advice, and always notes it’s not a lawyer or financial advisor. Tone guidelines emphasize warmth and clarity, with minimal formatting and bullets only when absolutely needed.