Overview
- Anthropic returned the high-end Claude Fable 5 to users in early July with strict per-user limits and strengthened safety classifiers that block or reroute some queries.
- Developers and a benchmarking group presented logs and a billing screenshot showing many sessions billed as Fable were actually handled by the older Opus 4.8 model without explicit user notification.
- BridgeMind’s BridgeBench scores for Fable 5 plunged because intercepted tasks were scored as zero after the classifier routed them to Opus 4.8, causing large drops in debugging, refactoring, and hallucination metrics.
- Security researchers publicly demonstrated multiple jailbreaks of Fable 5 even after its restoration, and Anthropic launched a HackerOne 'Cyber Jailbreak' disclosure program as a formal response.
- The changes have provoked developer outrage over transparency and billing, and Anthropic says it will temporarily remove Fable from subscription plans after July 7 until capacity and its safety posture meet its standards.