Opus 4.8 ships mid-conversation system messages that preserve prompt cache hits
Anthropic shipped Opus 4.8 on May 28 with four additions — mid-conversation system messages that hold cache hits when instructions change at runtime, adaptive thinking that fires reasoning only when needed, refusal categories in stop_details, and task budget support — while Managed Agents on AWS reached GA the next day with IAM-scoped policies and self-hosted sandboxes. Teams building multi-turn agents can now swap instructions mid-run at cache prices and run full multiagent orchestration inside their own AWS accounts today.
Source: platform.claude.com ↗
you can send role: "system" messages at non-first positions in the messages array, preserving prompt cache hits when instructions change during a long-running session.
Anthropic release notes
Why this matters
- → Mid-conversation system messages let you swap instructions mid-run without losing prompt cache savings, cutting costs for multi-turn agents.
- → Adaptive thinking reduces wasted tokens by reasoning only when needed, improving efficiency at the same capability level.
- → Refusal categories let apps route different decline types to appropriate handlers, enabling smarter error handling.
Cache-aware mid-run instructions