What changed
A new "thread" parameter on /v1/messages enables server-side conversation state; the first request sends the full conversation with "thread": {"type": "create"}, and each subsequent turn sends only new messages with "thread": {"type": "continue", "previous_message_id": "<prev_id>"} โ the API reconstructs the full context before calling the model.
TL;DR
Anthropic's Messages API now supports server-side conversation state in beta (message-threads-2026-08-12 header required); turns after the first send only new messages, reducing per-turn payload size and keeping TTFT low and steady as a conversation grows โ most useful in long agent loops.
Developer signal
The SDK thread helper (client.beta.messages.thread(...) in Python/TypeScript; similar in C#, Go, Java, PHP, Ruby) is the recommended path โ it adds the beta header automatically, sends only new messages on each turn, and replays the full conversation in a new create request when the thread expires (thread_not_found 404). For manual use, send anthropic-beta: message-threads-2026-08-12 on every request of the thread. Critical rules: system and tools go on the create request only โ sending either on a continue request returns a 400 error. model, thinking, tool_choice, mcp_servers, and the top-level cache_control are "fixed fields" โ they must be the same on every turn, or you get thread_fingerprint_mismatch. cache_control is capped at 4 breakpoints per request when using threads. Your application must maintain its own copy of the full conversation: the API does not guarantee it stored a given turn, and you find out only on the next continue request (or on a 404). Threads expire 5 minutes after the last request and are deleted โค1 hour after last use. Not ZDR eligible, not HIPAA eligible. Does not work with Message Batches, token counting, compaction, context editing (except clear_thinking_20251015 with keep: "all"), or Inference Hooks.