September 24, 2026
Claude charges for some refusals: the rule applies in 3 categories
On Sep 24, Anthropic started charging again for a block that happens before Claude responds, in three categories. The company says 99.7% of Claude Code, Claude.ai and Cowork accounts did not run into such blocks. In the API, a billed refusal is charged at the rate of the model that was run.
ClaudeDevs
@claudedevs
Starting today we are again charging for requests that our defenses block before Claude responds. This only covers categories with a low share of false blocks: biology, distillation attacks and building the smartest LLMs. Over the past weeks we have seen several coordinated attacks on our systems, and this is one of the layers of defense. In recent tests, 99.7% of accounts using Claude Code, Claude.ai or Cowork did not hit a single one of these new "billed blocks". We tuned the classifiers behind the blocks we are charging for again as of today to a false positive rate below 0.1%. We know that is not zero, and we will keep improving them so they interrupt your work less often. If you think a request was blocked by mistake, send feedback through /feedback in Claude Code. https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback#how-refusals-are-billed
· 779.9K views
A refusal before the first token in these three categories used to be free. Now any refusal spends the request limit, and a block in the three categories also costs money.
In Claude Code you can report a wrongful block through /feedback. The Claude API beta lets you turn on an automatic retry of a refused request on another model, but Message Batches and partner platforms do not have that option.
Anthropic promises to keep cutting the number of wrongful blocks.
