Breaking

Anthropic Resumes Billing for Some Claude Safeguard Blocks Before Output

September 24, 2026 at 13:12 EDT

Security
Foundation Models

Anthropic is resuming charges for Claude requests that safeguards stop before the model produces an answer, but only for blocks or fallbacks triggered by biology, distillation, or frontier LLM development classifiers. The September 24 change narrows billing to categories Anthropic says have low false-positive rates, while requests blocked before output in other categories remain uncharged.

September 24, 2026 · Anthropic

Blocked requests get a price tag

Anthropic will bill no-output Claude safeguard blocks only in biology, distillation, and frontier LLM development, a narrow rule tied to coordinated abuse defenses.

charged safeguard categories

other pre-output category charges

151M

exchanges in the largest cited distillation campaign

Where the billing switch happens

Request sent

→

Classifier checks context

→

Block or fallback before output

→

Bill only if category matches

Recent precision gains

Recent safeguard updates reduced benign or defensive interventions before the billing change.

85%

fewer benign biology flags

60%

fewer cyber interventions per session

Billing boundary

Charged: biology, distillation, and frontier LLM development blocks before output.
Not charged: pre-output blocks in other categories.
Separate charge: fallback responses billed at the responding model rate.

Operational case

The rule raises the cost of repeated classifier probes and coordinated abuse.

Developer concern

False positives can interrupt long sessions and now matter for bills in the named categories.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.