Confirmed 2 Nov 2026 · Labs A one-minute read
Anthropic will hold your agent's conversation on its servers to cut waiting time.
Message Threads entered beta on November 2. The conversation sits in encrypted memory for 5 minutes, and customers on zero data retention or health-care terms are barred from it.
Each request carries its own conversation, so the lab stores nothing between calls.
Speed now has a price in retention, and the strictest customers cannot pay it.
Why it matters
Agents re-upload the whole conversation every turn, and that is most of the waiting. Holding it on the server fixes the latency and moves the state. The customers who most need long agent runs, in hospitals and in firms on strict retention terms, are the ones expressly shut out.
Established
- Anthropic's release notes date Message Threads to November 2, in beta.
- A thread expires 5 minutes after its last request, deleted within 1 hour.
- Not eligible for zero data retention, and barred under health-care readiness.
Not established
- The size of the speedup. Anthropic's own pages publish no latency figure.
- That it is new work. The beta header is dated August 12, 2026.
- That data leaves memory. Anthropic says threads never reach durable storage.
The five-minute readThe record, the timeline, the weight and 2 sources
The record
Anthropic's platform release notes carry a single entry dated November 2, 2026: long conversations and agent loops on the Messages API can now send smaller requests and get a lower time to first token with message threads, in beta behind the beta header message-threads-2026-08-12. Read at platform.claude.com, along with the feature documentation. How it works: the caller sends the full conversation, system prompt and tools once with thread set to create, then sends only new messages with thread set to continue and the previous response's id. The API keeps the conversation server-side, so each later turn uploads and re-tokenises only new content. Lifetime: a thread expires 5 minutes after its last request, each stored response restarts the 5-minute timer, and the API deletes a thread no more than 1 hour after its last use. Threads are best effort: if the API cannot continue one, a continue request returns a 404 thread_not_found error and the caller must replay the whole conversation in a new create request. Storage and eligibility: threads are held in memory, encrypted under a server-managed key, and the documentation states they are never written to durable storage. They are not eligible for zero data retention. Organisations with health-care readiness enabled cannot use them at all. Scope: the feature is in beta on every model, available on the Claude API, Claude Platform on Amazon Web Services, Claude models on Google Cloud under Anthropic management, and Claude in Microsoft Foundry. The documentation's examples use claude-opus-5-5. Pricing and rate limits are unchanged, and prompt caching is not automatic: callers still set cache_control themselves, with at most 4 breakpoints per request. Unsupported alongside threads: Message Batches, token counting, compaction on demand, context editing except one thinking-clear setting, and organisations with inference hooks turned on. Threads are append-only, so earlier turns cannot be edited, and fixed fields such as model and tool settings must stay identical on continue requests or the call fails with thread_fingerprint_mismatch. There is no interface to list, read or delete a thread. Neither page publishes a latency figure or a date for the beta's start beyond the header name. Not established: adoption, the measured speedup, and why the header carries an August date.
How it got here
- Aug 12 2026
The date carried in the beta header name, message-threads-2026-08-12, which suggests the interface was fixed months before this public entry. Anthropic does not explain the gap.
- Oct 6
Anthropic says it will lift cyber blocks for verified security teams that let it keep logs, trading retention for access. Edition 044.
- Oct 21
Anthropic launches an agent that keeps working on its own computer after the user closes the app. Edition 058.
- Nov 2
Message Threads enters public beta, holding conversations in encrypted server memory for 5 minutes, excluded from zero data retention and health-care workflows.
- Nov 2
Anthropic tells the Securities and Exchange Commission it may imminently meet its own definition of dramatically accelerated progress attributable to AI. Edition 070.
Enterprise buyers who signed zero retention terms and now watch the fast path exclude them; hospitals and insurers under health-care readiness, locked out entirely; privacy officers who must re-approve a vendor that holds state; competitors whose own interfaces keep conversations client-side; and agent builders deciding whether 5 minutes is long enough to design around.
What to watch
- Would strengthen itAnthropic publishes a measured latency or request-size improvement for message threads, or moves the feature out of beta, by June 30, 2027.
- Would weaken itMessage Threads is withdrawn, or its documentation still shows no latency figure and no general availability, by March 31, 2027.
- Would strengthen itAnthropic extends server-held thread lifetime beyond 1 hour, or makes threads readable and listable through an interface, by December 31, 2027.
Why it weighs this much
Cost is low. This is engineering already built, and the serving memory it consumes is small next to the compute it saves by not re-tokenising. Commitment is medium: documented eligibility rules and a named beta header are a public promise about where customer text sits, and a lab that has sold zero retention hard does not write such an exclusion lightly. Reversibility is high, because a beta can be withdrawn and the documentation already warns that threads are best effort and may vanish mid-conversation. Persistence is a quarter: 5-minute threads deleted within an hour leave nothing behind. We score it low and say why: this is plumbing, and its interest is the trade it names, not its weight.
The words
- Time to first token
- How long a user waits before any answer appears. On long agent runs most of it is spent uploading and re-reading the conversation.
- Zero data retention
- A contract term under which the provider keeps nothing from a request after answering. Enterprise and regulated buyers often require it.
- Beta header
- A dated string a caller must send to switch on an unfinished feature. The date in the name marks the interface version, not the launch.
- Append-only
- Able to be added to but not changed. Earlier turns in a thread cannot be edited or removed, only continued.
Sources
Your call
Which explanation do you believe?
A call is final. Other people's calls show after yours.