AWS bedrock: Anthropic model card: refusal billing by stop_details.category
Summary
Replaces the short note on refused-request billing with a detailed table of stop_details.category values (cyber, bio, frontier_llm, reasoning_extraction, general_harms, null) and explains when a refusal is billed before any output, plus mid-stream billing and quota behavior.
Security assessment
The change is purely about billing semantics for content-classifier refusals; the mention of 'cyber harm' describes a refusal category, not a security control, vulnerability, or hardening guidance. No auth, encryption, credential, or exposure content is added.
Evidence
+`"cyber"` | The request could enable cyber harm, such as malware or exploit development. Benign cybersecurity work can also trigger this category. | No
Diff
diff --git a/bedrock/latest/userguide/model-cards-anthropic.md b/bedrock/latest/userguide/model-cards-anthropic.md index 72d184a..805adc4 100644 --- a//bedrock/latest/userguide/model-cards-anthropic.md +++ b//bedrock/latest/userguide/model-cards-anthropic.md @@ -39 +39,16 @@ The following billing behaviors apply specifically to Anthropic models on Amazon - * **Refused requests (all Anthropic models on`bedrock-mantle`):** When a request is refused by a content classifier before inference begins, no input or output tokens are billed. Mid-stream refusals — where the classifier fires after output has begun streaming — are billed for tokens consumed before the block. Both refusal types return `stop_reason: "refusal"` in the response. + * **Refused requests** – A refused request returns HTTP 200 with `stop_reason: "refusal"`. Whether it's billed depends on when the refusal occurred and on `stop_details.category`. These rules apply on both the `bedrock-runtime` and `bedrock-mantle` endpoints. + +The following table describes each `stop_details.category` value and whether Amazon Bedrock bills a refusal in that category before any output. + +**`stop_details.category`** | **What it means** | **Billed before any output** +---|---|--- +`"cyber"` | The request could enable cyber harm, such as malware or exploit development. Benign cybersecurity work can also trigger this category. | No +`"bio"` | The request could enable biological harm, such as dangerous lab methods. Beneficial life sciences work can also trigger this category. | Yes +`"frontier_llm"` | The request could help with the development of competing AI models, which is restricted under Anthropic's commercial terms. Benign machine learning work can also trigger this category. | Yes +`"reasoning_extraction"` | The request asks the model to reproduce its internal reasoning in the response text. To get reasoning in a structured form instead, use adaptive thinking. This category has no recommended fallback model, so change the prompt instead of retrying on another model. | Yes +`"general_harms"` | The request falls under a usage-policy area outside the four named categories. Benign work can also trigger this category. | No +`null` | The refusal doesn't map to a named category. `null` is a valid final value, not a placeholder. | No + +Amazon Bedrock bills a refusal before any output like any other request, at the refused model's rates. For a refusal that arrives mid-stream, after partial output, Amazon Bedrock bills the input tokens and the output tokens already streamed, in any category. Treat any partial output as incomplete and discard it. + +In every case, the response's `usage` still reports token counts, and the request counts against your quotas. The set of billed categories can change.