What changed

Anthropic made mid-conversation system messages available without a beta header on Claude Fable 5, Mythos 5 and Opus 4.8 through the Claude API, Amazon Bedrock and Google Cloud.

Why it matters

Applications can change trusted instructions during a session instead of reconstructing the whole conversation, which improves orchestration but creates a new privilege boundary inside the message history.

What to do

Add explicit provenance to every system-message insertion, deny untrusted tools and retrieved text from writing that role, and replay long conversations across all three providers to verify ordering, caching and audit logs.

Caveat

Anthropic establishes provider availability; it does not prove identical behavior across hosts or immunity to instruction conflicts. Model support and provider serialization can change.