Message threads in beta: send only new messages on long agent loops

The Messages API can now keep a conversation on the server for a short time, so each turn sends only new messages. It affects teams running long agent loops.

what changed

On 2 November 2026 Anthropic added message threads to the Messages API, in beta. You turn them on with the message-threads-2026-08-12 beta header.

Today, every turn uploads the whole conversation again, and the API checks and tokenizes all of it again. With a thread, you send the full conversation once, with your system prompt and tools. After that, you send only the new messages and the id of the previous response. The API rebuilds the conversation on its side.

Some limits to know. A thread expires 5 minutes after its last request. The API deletes it no more than 1 hour after its last use. If a thread can no longer continue, a continue request returns a 404 error, and you send the full conversation again. The docs say a thread’s contents are held in memory, encrypted, and never written to durable storage. The feature is not eligible for zero data retention. Threads are in beta on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry.

why it matters

Long agent loops pay for their history on every turn. Requests get bigger and time to first token grows with the conversation. Threads aim to keep both lower and steadier. Teams with agents that run for dozens of turns, such as support or coding agents, will feel it most.

Threads are best effort, so your application must stay the source of truth. If you rely on the server copy alone, a gap of over 5 minutes breaks the loop. The zero data retention note also matters for regulated workloads.

what we would do

  • Keep your own full copy of every message, and build the replay path first. Treat the thread as a speed-up, never as storage.
  • Test the 404 path on purpose: wait out the 5 minutes and check that the agent recovers without losing a turn.
  • Ask your data owners before you use threads on any workload that needs zero data retention.
  • Measure time to first token and request size on one long-running agent before and after. Roll out only where the gain is clear.

sources

Drafted with Claude from the linked sources, reviewed by us.