Anthropic Says It Will Charge for Some Claude Requests Blocked Before an Answer
The policy covers three safeguard categories. Anthropic says mistaken blocks should be rare, but has not specified how it would handle a disputed charge.
Listen to this story
The audio brief
Story brief
3 key pointsAnthropic’s September 24 policy puts a price on certain requests its safeguards stop before Claude responds, limited to biology, distillation attacks and frontier-model development—not every request in those areas. The company says 99.7% of accounts in recent testing encountered none of the newly billable blocks, while the classifiers’ false-positive rate is tuned below 0.1%; neither figure rules out an individual...
- 01
The change applies only to pre-answer blocks in three specified categories; it does not charge for an answer users dislike.
- 02
Anthropic directs Claude Code users to report suspected mistaken blocks through /feedback.
- 03
The policy does not explain how users would receive a financial remedy for an incorrect charge.
Anthropic said September 24 that it would resume charging for some Claude requests stopped by safeguards before Claude answers. The company cited coordinated attacks on its systems in recent weeks. Billing for blocked attempts could make repeated probing more costly, but the policy also means a mistaken block could carry a charge.
Which blocks would cost money
The change applies when a safeguard blocks a request before Claude responds, not whenever someone dislikes an answer. Anthropic limits the new charges to three categories: biology, distillation attacks and frontier LLM development.
These are categories of blocked requests, not a charge on every request about biology or AI development. Anthropic calls the policy one layer of defense against the attacks it says it has seen. Charging for a rejected attempt changes its cost, not whether the safeguard lets it through.
Anthropic says this share of Claude Code, Claude.ai and Cowork accounts encountered none of the newly billable blocks in recent testing.
Anthropic says the classifiers behind those blocks are tuned to a false-positive rate below this level.
Rare is not the same as impossible
The two figures measure different things. The account figure describes how often people in Anthropic’s recent testing avoided these blocks altogether. It does not say whether blocks seen by other accounts were correct. The false-positive figure describes how the classifiers are tuned, not the share of customers who will receive a mistaken charge.
Anthropic acknowledges that its false-positive rate is not zero and says it will keep improving the classifiers to interrupt work less often. A low overall rate cannot settle whether any particular block was justified. For a user whose legitimate request is stopped, the distinction now affects both the work and the bill.
How to flag a suspected mistake
Anthropic directs Claude Code users who think a request was blocked incorrectly to report it with /feedback. The announcement does not specify how a charge for a mistaken block would be handled. Feedback offers a way to flag the interruption, but the financial remedy remains unclear.
The September 24 post said charging would resume “today,” without giving a time for the switch. It announces a billing change; it does not confirm that charges had already appeared. For affected users, the first bills will show how the stated policy is applied to requests that produce no answer.
Sources
- x.comClaudeDevs (@ClaudeDevs) on X
Reader comments
Newest comments first. Replies stay oldest first.