Optimize Safety Refusals Using Automatic Model Fallbacks
Manage safety refusals automatically with the new fallback parameter.
Configure your workflow to use 'fallbacks' so that if Fable 5's safety classifiers block a request, a secondary model automatically takes over without manual intervention. 🛡️
The Scenario
You are working on sensitive research (like cybersecurity or biology) and your prompts are being flagged by Claude Fable 5's new, stricter classifiers.
Before & after
If a model refused a prompt for safety reasons, you would manually rewrite the prompt three or four times to guess which word triggered the block, wasting 10–15 minutes.
Check the 'stop_reason' or 'stop_details' in the API or interface. If it says 'refusal', use the 'fallbacks' parameter to automatically re-run the prompt on a different model in seconds.
The Prompt
[SYSTEM_INSTRUCTION: If using API, set "fallbacks": ["claude-3-5-sonnet-latest"]]. I am requesting a complex analysis of [TOPIC]. If this request triggers a safety refusal on Fable 5, please automatically fallback to a secondary model to provide the non-violating components of the data.
Claude Fable 5 introduces sophisticated safety classifiers. If a request is refused (stop_reason: 'refusal'), you are not billed for the generation. You can now use an opt-in 'fallbacks' feature to automatically route these blocked queries to a different model.
Source
Claude Platform - Claude API Docs"An opt-in fallbacks parameter... re-runs refused requests on another model, billed at the fallback model's rates."