Claude API message threads beta: send only the new messages
Claude API message threads keep the conversation server-side so agent loops send only new messages. Threads expire in 5 minutes, so keep your own history.
Every turn of a Claude agent loop has resent the whole conversation. Turn 40 ships turns 1 through 39 again. As of November 2, Anthropic has a beta that stops that. Claude API message threads keep the conversation on Anthropic's side, so each request carries only what is new. The catch is in the expiry rules, and they decide whether you can lean on it.
What actually happened
Anthropic's release notes list message threads as a Messages API beta, turned on with the message-threads-2026-08-12 beta header. The API keeps the conversation server-side, and each later request sends only the new messages. Anthropic says the result is smaller requests and a lower time to first token for long conversations and agent loops.
Two numbers matter. A thread expires 5 minutes after its last request. It is deleted no more than 1 hour after its last use.
The beta runs on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry. Beta features follow Anthropic's usual beta header rules, so the behavior can change before it goes general.
Why it matters for your business
This is a speed fix for agents that work in bursts. A support agent that pulls an order, checks a policy, drafts a reply and calls three tools makes many fast requests in a row. Smaller payloads mean less upload time and a faster first token on each step. Customers waiting on a chat window feel that.
Five minutes is a short leash. A thread does not hold a conversation that waits for a human. If your customer goes to lunch, or your agent pauses for an approval, the thread is gone. So the server copy is a cache, not your record. Keep the full history in your own database and be ready to rebuild the request from it when a thread has expired.
Test the failure path first. Before you turn this on, find out what your code does when it references an expired thread. If the answer is "the agent loses its memory mid-task," fix that before you chase the latency win.
Mind the data story. Anthropic deletes a thread within an hour of last use. If your privacy policy or a customer contract says how long conversation data sits with vendors, read the feature docs and confirm it fits before you ship.
Key takeaways
- Message threads are in beta on the Messages API with the
message-threads-2026-08-12header - The API keeps the conversation server-side, so each request sends only new messages
- A thread expires 5 minutes after its last request and is deleted within 1 hour of last use
- Treat the thread as a cache and keep the full history in your own store
- Test what your agent does when a thread has expired before you turn it on
Slow agent, long conversations? We profile where your agent loop spends its time, add caching and threads where they fit, and keep the conversation record in a database you own. See how we build agent systems, or tell us where it lags.
Sources: Anthropic: Release notes, Anthropic: Beta headers.
- #anthropic
- #claude-api
- #message-threads
- #ai-agents
- #latency
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Trigger.dev security incident: rotate your job secrets now
Trigger.dev says an attacker reached environment variables for a subset of orgs on November 1. Second env-var breach in two weeks. Rotate and cap keys now.
Read itNorthflank breach: AI API keys abused, 9 dev platforms hit
Northflank says customers saw stolen AI API keys used after its Oct 30 breach. Wiz counts nine AI dev platforms hit in October. Rotate and cap every key.
Read it