Anthropic's message threads beta lets Claude agents stop resending the whole conversation

Developers add one thread parameter to a Messages request. The first request carries the full conversation, the system prompt and the tools with "thread": {"type": "create"}. Every later turn sends only the new messages and the previous_message_id of the last response, and leaves system and tools out, because the thread keeps them[2]. Requests need the message-threads-2026-08-12 beta header, and Anthropic's SDKs ship a thread helper that adds the header, builds the requests and replays the conversation automatically when a thread can't be continued[2].

Threads are deliberately short-lived. Anthropic holds them in memory, encrypted, without writing them to disk[1]. A thread expires 5 minutes after its last request and is deleted no more than an hour after its last use, and a continue request on a dead thread returns a 404 that tells the client to replay the full conversation[2]. Anthropic is explicit that the application stays the source of truth and should keep its own copy of every message[2].

Anthropic has already turned threads on by default for most Claude Code users, and says requests of 500k to 1M tokens saw median time to first token fall by about 35%[1]. It names long coding agents, computer-use agents that would otherwise re-upload screenshots each turn, and document- and image-heavy workflows as the best fits[1].

It saves time, not money. Claude still reads the whole conversation every turn, so token usage is unchanged, and standard prices and rate limits apply[1]. Prompt caching works as before, but developers place the breakpoints themselves, with a limit of four per threaded request[2].

The beta covers every Claude model on the Claude API, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed)[2]. Threads aren't eligible for zero data retention or HIPAA, and they don't support Message Batches, token counting, compaction on demand or most context editing[2]. Threads are also append-only: each response id can be continued once, and editing or branching an earlier turn means starting a new thread[2].

As agent sessions stretch into hundreds of thousands of tokens, the delay before the first token becomes a real product problem, and threads attack it at the API layer instead of leaving every agent framework to work around it. The Next Web notes the launch follows Anthropic cutting cache prices by 75% in September as long agent sessions push up AI bills, so cost and latency are now being tackled separately[3].