Anthropic Publishes Prompting Claude Opus 5.5 Guide, Warns Old Opus 5 Effort Settings Now Blow Out Tokens

Anthropic's new Opus 5.5 prompting guide (front-paged on HN Sept 28) documents nine behavioral shifts from Opus 5: default effort dropped from high to medium, thinking is always on and can't be disabled, and existing effort settings will run longer and cost more per turn. New refusal categories (biology, cybersecurity, reasoning_extraction) return stop_reason 'refusal' with fallback semantics. Anthropic also details a system-prompt recipe to stop unattended agents from ending turns with progress reports instead of tool calls.