Claude API adds message threads with retention limits

Anthropic put Claude API message threads into public beta on the Messages API on 2 November 2026, according to its documentation and launch post. An application sends the full conversation once, then submits only new messages along with the previous response identifier. For teams running long agent conversations, that changes how requests are built. It also creates a temporary server copy of the conversation, and Anthropic's documentation says that copy rules out zero data retention and HIPAA use.

The primary sources are Anthropic's message threads documentation, its launch post and the API release notes.

Without a thread, every Messages API call resends the full history, and the API validates and tokenizes all of it again. With a thread, the first request carries thread: {"type": "create"} plus the full conversation, system prompt and tools. Later calls use thread: {"type": "continue", "previous_message_id": "..."} and include only the new messages. Every request needs the message-threads-2026-08-12 beta header.

• Continue requests leave out the system prompt and tool definitions, and sending them again returns a 400 error.
• The model, thinking configuration, tool choice and other fixed fields must match the original request, or the call fails with a thread_fingerprint_mismatch error. To change the system prompt or tools, a developer starts a new thread with the full conversation.
• A thread expires five minutes after its last request. A continue call after that returns a 404, and the response does not say why.

This changes how requests are handled and leaves the application's conversation database in place. Claude API conversation state therefore still belongs to the application, which must keep its own complete record and be able to replay it as a new thread. Anthropic's SDK helper can replay automatically on some missing thread errors, but interrupted streams and unsupported requests still need handling in code.

Claude API message threads and the retention decision

Anthropic's documentation says message threads are not eligible for Claude API zero data retention or HIPAA, and organisations with HIPAA readiness enabled cannot use them. Anthropic says thread contents are encrypted in memory and never written to durable storage, and are deleted no more than one hour after last use. That does not make the feature zero data retention eligible. There is also no API to list, read or delete an individual thread.

The feature runs on more than one hosting route. Anthropic lists the Claude API, Claude Platform on AWS, Anthropic managed Claude Models on Google Cloud and Claude in Microsoft Foundry. The beta covers every model, with the same data handling limits. An operator should check those limits against its contracts and internal data policy before moving production conversations over.

Threads also exclude Message Batches and compaction on demand. An agent that depends on either cannot simply add the thread parameter to its existing calls.

Anthropic's launch post says median time to first token falls by about 35 percent on requests of 500,000 to 1 million tokens. That is a company figure, it applies only to very long sessions, and the sources reviewed for this brief do not show independent measurement. Many production agents run far below that size, so the gain for a typical workload may be smaller. The same post says message threads are on by default for most Claude Code users, which means Anthropic's heaviest agent traffic is the evidence base for its own claim.

On billing, Anthropic's documentation says threads do not change pricing or rate limits. The launch post adds that Claude still reads the full conversation on every turn, so token usage stays the same. Threads do not switch on Claude API prompt caching either, so a developer must set a top level cache_control field or place breakpoints, and a fifth caching breakpoint on a thread request returns a thread_unsupported_request error.

A fair test therefore compares request payload size, time to first token, cache reads and recovery after expiry on the same conversation. A smaller network payload is not proof of a smaller bill. Teams auditing their API calls ahead of the end of Claude Opus 4.8 fast mode on 30 November can run this test in the same pass. For wider context on what agent usage costs, see our comparison of coding agent pricing across Claude Code, Devin, Replit and GitHub Copilot.

Guillaume Lebedel, co founder and CTO of StackOne, an AI agent infrastructure company, commented on LinkedIn that he expects computer use agents to adopt threads before coding agents do. His reasoning, as reported by The New Stack, is that a screenshot at every step makes resending history especially heavy. That is one practitioner's expectation, not a measured result. Some developers on X questioned whether a five minute expiry is long enough. Anthropic says each successful request refreshes the window, so an agent making regular calls can keep one thread alive through a long task.

A few points are worth watching.

• The Claude message threads beta carries an August 2026 version date in its header. Anthropic's pages reviewed here give no general availability date.
• The five minute expiry and the one hour deletion limit apply on every request, so idle agents will hit the 404 path.
• Before adoption, confirm that contracts, HIPAA status and any zero data retention agreement allow a temporary server side copy.

The practical choice is to test the beta on eligible, long conversations while keeping a reliable stateless fallback. The five minute expiry can be handled in code. A data handling requirement the feature does not meet cannot. For where conversation state sits inside a business assistant, see our guide to building an internal AI assistant on a foundation model.