Claude Platform changelog 11 2: Messages API adds message threads, available in beta, requiring the message-threads-2026-08-12 beta header. The first request sends the full conversation, system, and tools as usual, with "thread": {"type": "create"} added; subsequent turns send only the new message and the ID of the previous response ("type": "continue"), with the server reassembling the full conversation before calling the model. Threads expire 5 minutes after the last request and are deleted within 1 hours at the latest. Available on Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic-hosted), and Claude on Microsoft Foundry.
Confidence: High (80) · Event time 2026-11-02
Rationale
- Without threads, the entire conversation must be re-uploaded on every turn, and the API must re-validate and tokenize it; for agent loops involving dozens or hundreds of tool calls, request bodies grow larger and time-to-first-token increases. Official documentation states that with threads, most work targets only the new turn, resulting in lower and more stable time-to-first-token.
- It does not replace prompt caching: cache breakpoints are still set by developers, and top-level cache_control can be used to automatically cache each turn; both must be used together to save both latency and cost.
Usage
- Add anthropic-beta: message-threads-2026-08-12 to the request headers; add thread: {type: "create"} in the first turn, then use thread: {type: "continue", previous_message_id: <ID of previous response>} for subsequent turns, sending only the new message and omitting system and tools (sending them again will trigger 400). The official SDK includes thread helper objects that automatically handle the beta header and replay after 404. See the “Message threads” page at platform.claude.com/docs for documentation.
Risks
- Threads are “best-effort”: in cases such as expiration, responses generated by fallback models, or changes in user or workspace, continue requests will return 404, requiring a new create with the full conversation; therefore, applications must still store all messages themselves and pre-implement replay logic.
- Fields such as model, thinking, and tool_choice must remain unchanged throughout the thread; changes will trigger thread_fingerprint_mismatch; modifying system or tools on models like Opus 5.5 will invalidate existing thinking blocks. Behavior may change during the beta phase.