Claude API Threads Cut Long-Agent Lag 35%

Anthropic on Nov. 2 opened a public beta of message threads, a Messages API feature that lets AI agents stop resending their entire conversation on every turn. The company says median time to first token falls about 35% for requests of 500,000 to 1 million tokens, while token usage and prices stay the same.

Key takeaways

• After a first full request, agents send only new messages plus the previous response ID.
• Anthropic reports a roughly 35% cut in median time to first token at 500,000 to 1 million tokens; billing does not change.
• Threads live in encrypted memory, expire five minutes after the last request and are not eligible for zero data retention.

How Claude API message threads work

Developers add a thread parameter to a Messages API call along with the beta header message-threads-2026-08-12. The first request carries the full conversation, the system prompt and tool definitions, and creates the thread.

Each later request sends only the new messages and the ID of the last response, using "type": "continue". Continue requests must leave out system and tools; including either returns an error, according to Anthropic's documentation.

Threads are append-only. Each response ID can be continued once, so an app that wants to edit history or branch a conversation has to start a new thread. The docs say the feature is in beta on every Claude model.

Faster responses, same token bill

The gain is in latency, not cost. "Claude still reads the full conversation on every turn, so token usage stays the same," Anthropic wrote in its announcement. Standard token rates and existing rate limits apply.

The company says the benefit grows with context size. A coding agent can make hundreds of requests in one session, and once a conversation reaches hundreds of thousands of tokens, re-uploading it adds delay before Claude starts replying. Message threads are already on by default for most Claude Code users, Anthropic said.

Expiry, retention and limits

A thread expires five minutes after its last request, and the API deletes it no more than one hour after last use. Between turns, the state is held in memory, encrypted under a server-managed key and never written to durable storage.

That design means apps must keep their own copy of the conversation. If a thread has expired, a continue call returns a 404 thread_not_found error and the client must replay the full history in a new create request. There is no API to list, read or delete a thread.

Threads are not eligible for zero data retention or HIPAA use. They do not work with Message Batches, token counting, on-demand compaction or most context editing, and they allow at most four cache_control breakpoints. Prompts are not cached automatically inside a thread.

Context: how it compares with OpenAI

OpenAI's Responses API offers a similar chaining feature through previous_response_id. Its documentation likewise says that all previous input tokens in the chain "are billed as input tokens." The difference is retention: OpenAI stores response objects for 30 days by default, while Anthropic's threads vanish within an hour.

The feature is the latest in a run of Claude Platform changes. Anthropic raised the Messages API request-size cap from 32 MB to 128 MB on Oct. 28, its release notes show, weeks after Claude Opus 5.5 arrived 20% cheaper.

What's next

Message threads are available on the Claude API, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed). Anthropic has not said when the beta will end. Teams should test whether the five-minute window fits pauses in their agent loops.

FAQ

Do Claude message threads lower API costs?

No. Claude still processes the whole conversation each turn, so input tokens are billed as before. The benefit is lower latency.

What happens when a thread expires?

The API returns a 404 thread_not_found error. The app resends the full conversation to create a new thread and continues from the new response ID.

Can I use message threads with zero data retention?

No. Anthropic's documentation lists threads as not eligible for zero data retention or HIPAA-ready organizations.

References

• Make long-running agents faster with message threads — Anthropic (Nov. 2, 2026)
• Message threads — Claude Platform Docs
• Claude Platform release notes — Anthropic
• Conversation state — OpenAI API docs