Politics and regulation·September 25, 2026, 10:21
Anthropic to bill for certain blocked Claude requests
AI-generated and checked against the sources listed below.
The AI company Anthropic is starting to charge for requests that safety filters stop before Claude responds, but only within three narrow areas. According to Anthropic, the change affects almost no ordinary users.

Anthropic has announced that from now on, the company will bill for certain Claude requests that are blocked by safety filters before the chatbot gets to respond. That is according to several media outlets that have covered the story.
It only concerns three specific areas: biological safety (for example, questions about biological risks), so-called distillation attacks (attempts to trick Claude into revealing how the model is trained) and the development of advanced language models. In these categories, a blocked request will therefore cost money even though Claude does not deliver a real answer.
According to Anthropic, the background is "coordinated attacks" on the system. The company's reasoning is that a free refusal is in practice a free chance to probe the model's limits. If being refused is free, people have an incentive to keep trying to find loopholes. By making it expensive to be blocked, Anthropic wants to make these kinds of repeated attacks less attractive.
Technically, this works by having a so-called classifier filter (a smaller model that reads the request before the Claude model itself does) assess whether the request falls within the three risk areas. If it does, the response is sent back as blocked, and the user is still charged for the tokens (units of computation) the request has used.
Anthropic says that 99.7 percent of accounts on Claude Code, Claude.ai and Cowork did not experience a single one of these paid blocks in testing. At the same time, the company acknowledges that the system is not perfect and that erroneous blocks of entirely legitimate requests may occur. Claude Code users can report a suspected erroneous block via the /feedback command, but there is not yet a published procedure for how an unjustified charge will be refunded.
For ordinary users, the change therefore means very little in practice, as it only affects a small number of specialized and potentially harmful requests. But it shows how AI companies are increasingly using pricing as a tool to combat misuse of their systems.
Sources
Get the week's AI news in your inbox
Choose your level, topics and length. One email a week, unsubscribe at any time.
Subscribe to Promptly Newsletter



