On 2 November 2026 Anthropic moved message threads for the Messages API to public beta. Send anthropic-beta: message-threads-2026-08-12 on every turn. Turn one adds thread: {type: create} with the full messages, system, and tools. Later turns send only new messages plus thread: {type: continue, previous_message_id} pointing at the previous response id, and omit system and tools. Response JSON stays a normal message object. Threads expire about five minutes after the last request and are deleted within about an hour. On continue failures, read error.details.error_code for thread_not_found before you replay the full conversation.
Dated 3 November 2026 from the release notes and Message threads documentation. Yesterday we covered ChatGPT Pro/Business API credits—org balances, not request shape. Ten days ago, 128 MB limits addressed payload size; threads address not re-uploading entire histories each agent loop turn.
Create vs continue JSON
Public beta still uses POST /v1/messages. The only thread-specific piece is the top-level thread object shown below.
{
"model": "claude-opus-5-5",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "Shipping question?"}],
"thread": {"type": "continue", "previous_message_id": "msg_01..."}
}
Including system or tools on continue—even as null—returns 400. Documented first-party and several managed platforms support threads; elsewhere keep sending full message arrays.
Which id goes in previous_message_id
Use the prior response’s top-level id, not a client-side message key. Continue from the latest response only; fork or create anew for branches. Responses do not confirm whether a turn was stored—you learn on the next continue or 404.
Fixed fields on every turn
model, thinking, top-level cache_control, and the beta header set must match create on every continue. Changes yield thread_fingerprint_mismatch. Threaded requests allow at most four cache breakpoints.
Five-minute TTL and 404 replay
State is in-memory and best-effort. thread_not_found means replay with a full create from your own transcript. Token counting requests cannot include thread—count the full conversation instead.
Tools, pause_turn, and streaming
Continue with tool_result user messages after tool_use. For pause_turn, continue with empty messages. Wait for message_stop before continuing a stream.
Validating thread JSON locally
Keep separate JSON Schema variants for create and continue. Use JSON validation to catch system/tools on continue turns.
FAQ
Same as Managed Agents session threads?
No. Message threads are short-lived Messages API state.
Skip the beta header on continue?
No. Every threaded request needs message-threads-2026-08-12.
Conflict with 128 MB?
No. Threads reduce repeat upload; byte limits still apply.
In-thread compaction?
Threads are append-only; start a new create with a summary message.
Need API version 2026-10-21?
Orthogonal to beta threads; follow your SDK and platform defaults.
Extra thread_id in responses?
No. Chain response ids only.
Summary
Message threads trade create/continue JSON discipline for smaller follow-up requests and steadier TTFT in long agents—keep your own transcript and handle replay.