The return of Claude Fable 5 on July 1st sparked immediate user complaints about broken performance. However, two major benchmarks tell different stories.
BridgeBench reported severe drops in coding metrics. Their debugging score fell from 86.2 to 25.9. The cause: Anthropic's new safety classifier intercepted 9 of 12 tasks, rerouting them to Claude Opus 4.8. BridgeBench scores these reroutes as zero.
The classifier was deployed to block a known jailbreak technique that could identify software vulnerabilities. It is now overly sensitive, flagging routine coding work as security risks.
Arena.AI's human-preference votes, conducted across thousands of prompts, found Fable 5's performance largely unchanged or even improved in categories like document analysis and creative writing. Performance dips in coding tasks align precisely where the classifier intervenes.
The model itself is not weaker. The gatekeeper is blocking it from doing its job in specific domains. Developers working near security-adjacent code will notice the impact. Writers and researchers will likely see no difference. Anthropic has acknowledged the classifiers are too broad but has given no timeline for refinement.