Anthropic adds Claude self-assessment for broad misuse monitoring

Anthropic introduced self-assessment, a post-response safety check where Claude evaluates a conversation for potential misuse, costing about 10 extra tokens and under 2% of inference cost. The company reports it flagged 78% of harmful conversations and 94% of high-priority cases with roughly a 1% false-positive rate.

• Assessment adds ~10 tokens and under 2% to inference cost
• Flagged 78% of harmful and 94% of high-priority conversations, ~1% false positives
• In August 2026 it surfaced 100,000+ child-exploitation flags from one API customer
• Runs on all Claude models since June 2026 and is not billed separately