On Nov. 2, 2026, Anthropic opened a public beta of message threads in the Claude Messages API. The feature targets long-running agents that today resend the full conversation on every turn. With threads, the first request carries the whole history and starts a server-side session; later turns send only new messages plus the previous response id, which Anthropic says cuts median time to first token on very large sessions.
Key facts
| Item | Detail |
|---|---|
| Announced | Nov. 2, 2026 |
| Status | Public beta |
| Beta header | message-threads-2026-08-12 |
| Anthropic's speed figure | Median TTFT down about 35% on 500k–1M token requests |
| Thread lifetime | Expires 5 minutes after last request |
| Token usage | Unchanged |
What Anthropic launched
Message threads is now available in public beta on the Claude Platform. Anthropic's Nov. 2 article states that agents no longer need to resend the full conversation history on every turn, reducing latency for long-running sessions. You add a thread parameter to Messages API requests and send the message-threads-2026-08-12 beta header on every request in the thread.
Anthropic highlights long-running coding agents, computer use agents, and document or image-heavy workflows, and notes that message threads is on by default for most Claude Code users.
How message threads work
The platform docs describe a three-step pattern:
- Create. Send the first request as you do today, with the full conversation,
system, andtools, and add"thread": {"type": "create"}. The API stores the conversation under the responseid. - Continue. Send only messages since the previous response, with
"thread": {"type": "continue", "previous_message_id": "..."}. Leavesystemandtoolsout; the thread keeps the values from the create request. - Replay. Threads are best-effort. A thread expires 5 minutes after its last request, and the API deletes it no more than 1 hour after its last use. If a continue hits an expired thread, the API returns a 404 and your client must resend the full conversation in a new create request. Your application stays the source of truth—keep your own copy of every message and build the replay path from the start.
Between turns, Anthropic holds each thread in memory, encrypted and not written to disk. The SDK thread helper (for example client.beta.messages.thread in Python) handles replay when a thread expires.
The speed claim and what it does not change
Anthropic's article gives a concrete latency figure with qualifiers: with message threads on by default for most Claude Code users, requests with 500k to 1M tokens see median TTFT drop by about 35%. That is a median improvement on very long requests, not a blanket speedup for every workload.
The docs are explicit that Claude still reads the full conversation on every turn, so token usage stays the same. Standard Claude token rates and existing rate limits apply. Prompt caching works as before; a thread request allows at most four cache_control breakpoints (the top-level cache_control counts as one).
Where it's available and data-handling limits
Anthropic lists message threads on the native Claude Platform, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry. Teams already running Claude on Google Cloud through Anthropic-managed endpoints can adopt the same API pattern described in Anthropic-Managed Claude Goes GA on Google Cloud.
Platform documentation marks message threads as not ZDR eligible and not HIPAA eligible. Thread contents live in encrypted memory only and are deleted no more than 1 hour after last use.
What developers should check before turning it on
From Anthropic's docs, a practical pre-flight list:
- Send
systemandtoolson the create request only; sending either on a continue request returns a 400 error. - Keep fixed fields identical on every request in a thread (
model,tool_choice,thinking, top-levelcache_control, and the sameanthropic-betavalues). A mismatch returnsthread_fingerprint_mismatch. - Include the
message-threads-2026-08-12beta header on every turn. - Plan for 404 replay: store the full conversation locally and resend it in a new create when a thread expires.
- Stay within four
cache_controlbreakpoints per thread request if you rely on prompt caching. - Skip threads if your organization has HIPAA readiness enabled.
What it means
Analysis (ours). Message threads cut upload and validation overhead, not model tokens or rates. Median TTFT gains on 500k–1M token agent loops could feel meaningful in coding and computer-use stacks, but the five-minute idle expiry means you still need durable transcripts and replay paths. Regulated teams must weigh the ZDR and HIPAA exclusions even though threads never hit disk.
What to watch
Thread lifetime policy, replay behavior after 404s, and whether more Anthropic products default threads on beyond Claude Code. Docs already warn that changing fallbacks mid-thread triggers thread_fingerprint_mismatch—worth tracking as teams wire server-side fallbacks into agent loops.