Anthropic to Charge Developers for Blocked Claude Prompts

Anthropic has announced plans to bill developers for API requests that are blocked by safety classifiers before any output is generated. The new policy targets three specific categories of refusals: biological safety risks, model-distillation attacks, and frontier LLM development. By charging for the input tokens of these blocked prompts, the company aims to establish a defensive financial barrier against coordinated automated probing, jailbreaks, and unauthorized model extraction.

According to Anthropic, the vast majority of users will not be affected by this change. In recent testing, 99.7% of accounts using Claude Code, Claude.ai, and Cowork did not trigger any of these newly billable blocks. Additionally, the company states that its safety classifiers are tuned to keep false-positive rates below 0.1%. However, because a blocked request still returns a standard HTTP 200 status code with a stop_reason of refusal, developers must actively monitor their API responses to identify when they are being billed for blocked inputs.

For practitioners, this shift means false positives will now incur direct financial costs alongside workflow interruptions. To mitigate this, developers can implement a server-side fallback by using the server-side-fallback-2026-07-01 beta header and setting the fallback parameter to default. This allows the system to automatically retry a refused request on a recommended alternative model. SDK middleware is also available to handle manual retries, which can utilize fallback credit tokens to avoid paying twice for prompt-cache writes.

To prepare for the billing change, developers should update their monitoring systems to inspect successful HTTP 200 responses for refusal indicators. Teams should log the stop_details.category, token usage, and fallback outcomes to reconcile future invoices. Those working in life sciences, synthetic-data generation, or advanced model evaluation face the highest exposure and should closely track these metrics, using tools like the /feedback command in Claude Code to report incorrect blocks.