Handle Claude Fable 5 cyber refusals in production — stop_reason, fallbacks to Opus, and when to use an uncensored API instead.
Claude Fable 5 adds safety classifiers that return stop_reason refusal with category cyber — on HTTP 200. Your dashboard thinks the call succeeded.
Anthropic documents that benign cybersecurity work can trigger this category.
Response shape
Branch on stop_reason, not content. stop_details.category may be cyber, bio, reasoning_extraction, or null. Empty content arrays are common.
Fallback options
Server-side fallbacks to Opus 4.8 (beta), SDK refusal-fallback middleware, or manual retry with fallback credit tokens.
When fallback is not enough
If you ship a security product to customers, depending on Opus fallback chains adds latency and cost. Dedicated uncensored inference (Icelake) keeps your SLA predictable.
Try uncensored AI free
No filters, no lecture, no training on your chats. Start in under a minute.
Treat classifier refusals as a first-class metric — not as rare edge cases.
FAQ
Are refusals billed?
Anthropic docs: refused requests with no output are typically not billed for generation; prompt cache costs may still apply without fallback credit.