Anthropic resumes billing for blocked requests in three categories

Anthropic has resumed charging for some requests that its safety classifiers decline before Claude writes anything, limited to three categories where it says false positives are rare. Its announcement names them as biology, distillation attacks and frontier LLM development, and presents the change as one layer of defense after what it calls coordinated attacks on its systems in recent weeks. Anthropic's documentation on refusals states the purpose as disrupting attempts to circumvent its safeguards at scale.

A declined request does not come back as an error. On Claude Fable 5.1, Fable 5, Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5, the API returns a normal successful response with empty content, a stop reason of "refusal" and a category naming the policy area. The documentation's billing table marks three categories as billed when the refusal arrives before any output: bio, frontier_llm, which covers help developing competing AI models, and reasoning_extraction, which covers requests asking the model to reproduce its internal reasoning in its reply. The page does not use the phrase distillation attacks. Billed refusals cost the same as any other request at the rates of the model that ran it, while early refusals in other categories, such as cyber, stay free. Every refusal counts against rate limits, and one that arrives mid-stream is billed for its input and the output streamed so far.

Anthropic's own figures suggest most people will not notice: it says 99.7% of accounts using Claude Code, Claude.ai or Cowork did not hit any of these billable blocks in recent testing, and that the classifiers are tuned to a false-positive rate below 0.1%, which it acknowledges is not zero. Developers have more to check. Because a refusal looks like a success, monitoring built on error rates never sees it, so the documentation advises counting refusals as their own signal. Applications that retry declined requests on another model should expect that a billed refusal is charged in addition to the retry. Anthropic says the billed list may change as it keeps measuring false positives, and asks anyone blocked incorrectly in Claude Code to report it with /feedback.