Fable 5 returns from shutdown with safety classifiers developers say now block ordinary code

Anthropic redeployed Fable 5 and Mythos 5 on July 1 after a 19-day period in which United States export controls forced the models offline. The company paired the return with a proposed industry framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners.

The technical change is narrow but consequential. A classifier layer covers cybersecurity, biology, chemistry, and model distillation, and any trigger can route a request to Opus 4.8. Only the cybersecurity classifier was retrained, and it is now more sensitive, which Anthropic acknowledges flags benign requests more often during routine coding and debugging. The company says the specific attack technique in an Amazon report is now blocked in more than 99 percent of cases, and that it will keep tuning the system to reduce false positives.

The user complaint is about who pays for those false positives. One widely shared post claimed a 321 dollar coding session in which classifiers routed most of the work to the more expensive Opus 4.8, with the cheaper Fable model handling a fraction of it. That figure comes from a single social account and is not independently verified, but the mechanism it describes, silent escalation to a pricier model when a classifier triggers, is confirmed by Anthropic's own documentation.