Requests involving distillation attacks, biology, and frontier LLM development that are rejected by safety filters will now trigger a charge for the user. Anthropic is applying the policy to Claude Code, Claude.ai, and Cowork to prevent attackers from probing its safeguards for free following coordinated attacks on its systems in recent weeks. Testing indicates 99.7% of accounts did not trigger these billable blocks, and the underlying classifiers are tuned to a false positive rate of less than 0.1%.

The company had previously paused charges for flagged requests around the launch of its Fable product. Anthropic will continue to refine its classifiers to reduce the frequency of interruptions for legitimate users while maintaining the blocks as a layer of defense against system abuse.