Claude’s Tier 3 Security Refusal Responses Now Charge Users: No Response Still Incurs Fees

Beating AI News Flash: Anthropic Resumes Billing for Three Categories of "Zero-Output Safety Rejections" Claude will now charge at its standard model rate even when it rejects requests before generating any output, if the rejection stems from biosecurity, frontier large language model (LLM) development, or "reasoning extraction"—a category referring specifically to attempts to elicit Claude’s internal reasoning processes. On June 2 this year, Anthropic had scrapped all charges for zero-output rejections, but is now reinstating billing exclusively for these three categories. The change only applies to these cases: other requests (such as those for cyberattacks or general harmful content) that get rejected before output remain free of charge. If Claude has already generated partial output before being blocked, it will always be billed normally. Anthropic noted that coordinated attacks targeting its system have emerged in recent weeks, so billing is being deployed as an additional defense measure. Official tests found that 99.7% of Claude Code, Claude.ai, and Cowork accounts did not encounter these three billing-triggering rejections, and the false positive rate for related safety classifiers is below 0.1%. However, false positive reports have already appeared on GitHub. One Claude Code user documented 23 false "reasoning extraction" judgments, including for routine tasks like podcast scripts and budget spreadsheets, with 7 cases terminating the current task. Another user reported that general software development tasks and Claude Code’s built-in workflow instructions also triggered similar rejections.