Anthropic reinstates charges for categories of Claude zero-output safety refusals

AT A GLANCE

The report says that Claude the three types of zero-output safety refusals have returned to standard-rate billing, and there are already reports of ordinary tasks being misclassified.

Article

Claude Charges reinstated for categories of zero-output safety refusals: you are charged even if not a single word is returned

Beating AI news flash: Anthropic has reinstated charges for categories of “zero-output safety refusals.”Claude If a request is refused before output due to biosafety, frontier large-model development, or “reasoning extraction,” it will be charged at the standard rate for the model actually run, even if not a single word is returned.

“Reasoning extraction” mainly refers to attempts to obtain Claude internal reasoning processes. Earlier this year, Anthropic June 2 had eliminated charges for all “zero-output refusals,” and is now reinstating charges for these categories.

This change applies only to these categories. Other requests, such as cyberattacks and general harm, will still not be charged if they are likewise refused before output; if Claude some output has already been generated before being blocked, standard charges have always applied.

Anthropic said that coordinated attacks against its systems have emerged in recent weeks, so it is also using charges as a line of defense. In official testing, 99.7%of Claude Code、Claude.ai and Cowork accounts did not encounter charge-triggering blocks in these categories, and the false-positive rate of the relevant safety classifiers was below 0.1%.

However, reports of false positives have already appeared on GitHub. A Claude Code user documented 23“reasoning extraction” false positives, including ordinary tasks such as podcast copy and budget spreadsheets, of which 7 directly terminated the task for that turn. Another user reported that ordinary software-development tasks and the workflow instructions built into Claude Code had also triggered similar blocks.

Original link https://m.theblockbeats.info/flash/368956

Key points

The report says that Claude when a request is refused before output due to biosafety, frontier large-model development, or “reasoning extraction,” it will be billed at the standard rate for the model actually run.
“Reasoning extraction” primarily refers to attempts to elicit Claude internal reasoning processes.
Reports said that Anthropic had, earlier this year,June 2 cancelled charges for all “zero-output refusals” and is now charging again for the above three categories.
Other requests, such as cyberattacks and general harm, remain free if refused before output; requests blocked after partial output will be charged normally.
Anthropic said coordinated attacks targeting its systems had emerged in recent weeks, so it is also using charges as a line of defense.
Official testing found that 99.7%'s Claude Code、Claude.ai and Cowork accounts did not encounter charge blocks in these three categories, and the relevant safety classifier's false-positive rate was below 0.1%。
Reports say GitHub already has reports of false positives; one Claude Code user recorded 23 instances of “reasoning extraction” misclassification, of which 7 directly terminated the task for that turn.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Put simply,Claude Even if no text is returned, model charges may still apply if a request is categorized as biosafety, frontier large-model development, or “reasoning extraction” and refused before output. Other types of zero-output refusals remain free of charge. The report also cites user feedback stating that podcast copy, budget spreadsheets, and ordinary software-development tasks have also been misclassified.

Why it matters to readers

Users may still be charged without receiving usable output; if an ordinary task is misclassified by the safety classifier, this may also result in both charges and task interruption.

Use Claude Code、Claude New users of.ai or Cowork should note: whether charges apply depends not only on whether there is text output, but also on which category of safety interception the system assigns to the request.

Risks and unknowns

The supplied materials do not include the original official Anthropic announcement, so the charging rules still require independent verification.
The materials do not specify the exact effective date for the resumption of charges, the appeal mechanism, or how the bills will be displayed.
The misclassification reports on GitHub are user reports; the materials do not state whether Anthropic has confirmed these cases.
“99.7%” and “below 0.1%” are officially reported test results relayed in news reports; the materials do not provide the test methodology.

Anthropic reinstates charges for categories of Claude zero-output safety refusals

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center