Redeploying Fable 5

Anthropic announced that its newest model, Claude Fable 5, will be available globally on July 1 after the U.S. government lifted export controls on June 30. The company had suspended access to both Fable 5 and Mythos 5 on June 12 because it could not verify user nationality in real time. Fable 5 will be offered on the Claude Platform, Claude.ai, Claude Code and Claude Cowork, with a 50% usage limit for Pro, Max, Team and select Enterprise plans until July 7, after which usage will be billed via credits. Anthropic also restored access to Mythos 5 for a set of U.S. organizations following approval on June 26 and is coordinating with the government to expand access to other domestic and international partners in the Glasswing program. The company released both models on June 9, with Fable 5 featuring stronger safeguards for general use and Mythos 5, with fewer safeguards, limited to trusted Project Glasswing partners for defensive cybersecurity. The export control directive followed a report by Amazon researchers that identified a method to bypass Fable 5’s safeguards, enabling the model to produce code that demonstrated exploitation of software vulnerabilities. Anthropic worked with the government and partners, including Amazon, to review the report and evidence. Testing showed that less capable models could also identify the same vulnerabilities and produce similar exploit demonstrations. In response, Anthropic trained an improved safety classifier that blocks the reported bypass in over 99% of cases, notifying users when a request is blocked and redirecting it to Opus 4.8. The new classifier increases false positives during routine coding and debugging tasks, and Anthropic plans to refine it to better distinguish legitimate requests. Anthropic emphasized that Mythos 5’s capabilities make it attractive to malicious actors, whereas Fable 5 was launched with the strongest safeguards ever applied, employing a defense‑in‑depth approach that includes multiple safety mechanisms and classifiers. The company acknowledges that no model can be fully impervious to jailbreaks, but it expects to find and fix major jailbreaks before malicious actors can exploit them. Anthropic is also proposing an industry‑wide framework for scoring jailbreak severity with Amazon, Microsoft, Google and other Glasswing partners.