THE ESSENTIALS IN 30 SECONDS

  • With a thread, developers send the full conversation once, then only new messages plus the previous response ID; system prompt and tools are kept from the first request.
  • Anthropic reports roughly 35% lower time to first token for requests of 500,000 to one million tokens, according to The Decoder.
  • Threads expire five minutes after the last request, do not change pricing or rate limits, and are not eligible for zero data retention.

Anthropic has released Message Threads, a public beta feature of its Messages API aimed at long-running agents. Until now, every API call had to carry the entire conversation history, which the service then validated and tokenized again on each turn. A thread lets the server hold that history briefly so that each new request contains only what changed.

How threads work

According to Anthropic's documentation, a developer opens a thread by sending the first request as usual, with the full conversation, system prompt and tools, plus a new thread parameter set to create. Later turns send only the new messages and the ID of the previous response. The system prompt and tools are not resent; the thread keeps the values from the first call, and certain settings such as the model must stay identical or the API returns an error. Using the feature requires the beta header message-threads-2026-08-12, and Anthropic's SDKs include a helper that builds these requests.

Anthropic says the main benefit is latency. As reported by The Decoder, the company measured time to first token falling by about 35% for requests between 500,000 and one million tokens. The documentation describes the gain more generally: requests get smaller and response start times stay lower and steadier as a conversation grows. Token usage itself does not shrink, because the model still reads the whole history, and the docs state that threads change neither pricing nor rate limits. Prompt caching still has to be requested separately.

Limits and data handling

Threads are best-effort. One expires five minutes after its last request, and if it can no longer be continued the API returns a 404 error, so developers must keep their own copy of the conversation and resend it in full. The documentation tells developers to build that replay path from the start.

The feature does not work with Message Batches, token counting, compaction or most context editing, and is unavailable to organizations with HIPAA readiness enabled. Anthropic says thread contents are held in memory under a server-managed encryption key, never written to durable storage and deleted within an hour of last use. Even so, threads are not eligible for zero data retention, which may matter for customers with strict retention agreements.

The beta covers a long list of Claude models and runs on Anthropic's own API as well as on AWS, Google Cloud and Microsoft Foundry. The Decoder names coding agents, computer-use agents and multi-document workflows as the intended use cases.