📝 TL;DR 📝 Claude Fable 5 returned globally on July 1 after the US export controls that forced it offline on June 12 were lifted. It comes back with a more aggressive cybersecurity classifier, one that will flag a slightly higher share of harmless coding requests as a deliberate trade-off for blocking more genuine misuse. If a normal request gets flagged, use the feedback button rather than assuming something is wrong with your prompt. Paid plans get Fable 5 included up to 50% of weekly usage through July 7, then via usage credits. 🧠 Overview 🧠 The most powerful Claude model got pulled offline for nearly three weeks after a US government export control directive, and it is now back with real changes underneath the hood. Understanding what actually happened, and why the new version behaves slightly differently, is useful both practically (you might hit a refusal on something ordinary) and as a window into how AI safety actually gets tuned in response to real-world findings rather than theoretical ones. Anthropic's own explanation of what happened is unusually detailed and worth understanding directly, because it clarifies something that got muddled during the outage: the vulnerability behind the shutdown was not unique to Fable 5 at all. 📜 The Announcement 📜 As of July 1, Fable 5 is available globally across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans, it is included for up to 50% of weekly usage limits through July 7, after which access shifts to usage credits. Mythos 5, the less-restricted sibling model, has been restored only to a specific set of US organizations approved by the government on June 26, with broader Project Glasswing access still being negotiated. Anthropic's account of the underlying trigger is notable: Amazon researchers had found a way to prompt Fable 5 into identifying software vulnerabilities and, in one case, producing code demonstrating how to exploit one of them. But Anthropic's own testing found that less capable models, including Opus 4.8, GPT-5.5, and Kimi K2.7, could identify the exact same vulnerabilities.