Anthropic redeploys Fable 5 with a new jailbreak classifier

Category: AI
Status: confirmed
Sources: Anthropic, CNBC, HN discussion
Summary: Anthropic began restoring Claude Fable 5 globally on 2026-07-01 after the US lifted the export controls imposed 2026-06-12. The redeployment ships a new safety classifier that Anthropic says blocks the reported jailbreak technique in over 99 percent of cases; the jailbreak, reported by Amazon researchers, had prompted the model to identify and in one case show how to exploit a software vulnerability. Anthropic says it is drafting a cross-industry jailbreak-severity framework with Amazon, Microsoft, Google, and other partners. Mythos 5, the same underlying model with fewer restrictions, is being restored to approved US organizations.
Comments: HN commenters flagged the usage terms: through 2026-07-07 a plan can spend up to 50 percent of its weekly limit on Fable 5, then continue on usage credits, and Fable 5 draws down usage faster than Opus 4.8.
Why it matters: The episode sets a template for classifier-gated redeployment and government pre-release review of frontier models, and the industry jailbreak-severity framework may standardize how labs triage misuse reports.
Follow-up: Watch for the published jailbreak-severity framework, independent testing of the new classifier, and whether Mythos 5 access widens.