Claude's Messages API adds message threads to cut resent history in agent loops

Anthropic launched message threads in public beta, letting agents send only new messages each turn to lower time to first token in long sessions.

Anthropic put message threads into public beta on the Messages API on Nov 2, 2026, per its announcement. Instead of resending the full history every turn, an agent sends the whole conversation once, then only new messages plus the previous response ID. If you run long coding or computer-use agents, that means less upload before Claude starts responding. Anthropic says that with threads on by default for most Claude Code users, requests of 500k to 1M tokens see median time to first token drop by about 35%. That is Anthropic's own Claude Code figure, not an independent benchmark. AlphaSignal reports that median request size in long Claude Code sessions fell from about 800 KB to 1.8 KB.

It saves time, not money. Claude still reads the full conversation each turn, so token usage, pricing and rate limits are unchanged. The docs say nothing is cached unless you add a top-level cache_control or your own cache breakpoints.

To use it, you send the message-threads-2026-08-12 beta header on every request. The first request uses "thread": {"type": "create"}. Later requests use "type": "continue" with previous_message_id. Leave out system and tools on continue requests; sending either, even as null, returns a 400 error, per the docs.

The limits matter for production. Threads are best-effort, expire 5 minutes after the last request, and return a 404 when they can't be continued. The docs tell you to keep your own copy of every message and replay it. TNW notes that threads aren't eligible for zero data retention or HIPAA setups. Unite.AI adds that threads don't support Message Batches or token counting.

One developer's write-up points out a practical risk. An agent that waits on a slow test suite or on human approval can come back to an expired thread. Build the replay path first.

Sources