- Published on
Claude API Gets Message Threads Beta β Long Conversations 35% Faster
The longer an AI agent loop runs, the slower it gets. Anthropic attacked that problem at the root.
On November 2, 2026, Anthropic added a new beta feature to the Claude Messages API: Message Threads. It lets long conversations and agent loops send smaller requests and receive a faster time to first token by keeping conversation context server-side.
The result: roughly 35% lower median time to first token (TTFT) on long contexts.
Why Do Long Conversations Get Slow?
Here is how the standard Claude API works. After ten turns of conversation, the eleventh request must include all ten previous turns. The model needs the full history to understand context.
In agent loops, this compounds quickly. When a conversation grows to hundreds of thousands of tokens, every single request transmits that entire payload over the network. The larger the data, the longer the wait for the first response byte.
How Message Threads Work Differently
Message Threads keeps conversation context on Anthropic''s servers.
The first request still carries the full conversation. But from the second request onward, the client sends only the previous response ID plus the new message. The server already holds the context, so the client sends only what is new.
| Request | Standard API | Message Threads |
|---|---|---|
| 2nd request | Full conversation | Previous ID + new message |
| 10th request | Accumulated full history | Still just new message |
| TTFT pattern | Grows with conversation length | Stays consistent |
Anthropic''s own measurement shows a 35% reduction in median TTFT for conversations between 500,000 and 1 million tokens, based on Claude Code request data. One external developer reported roughly 30% improvement on a 600K-token agent loop.
Does It Cut Token Costs?
No. This is an important distinction.
Message Threads reduces network payload size, not the number of tokens the model processes. The server still reads the full context to generate each response. Billing continues to apply to the full input token count.
Faster, but not cheaper.
How to Activate It
Message Threads is activated by a single API header:
anthropic-beta: message-threads-2026-08-12
After the first request, Anthropic''s server returns a response ID. Pass that ID in the previous_message_id field of the next request.
Your application still needs to maintain its own full conversation transcript locally. Server-side threads expire five minutes after the last request. When a thread expires, the API returns a 404, and your app can restart a new thread using its locally stored history.
Where Is It Available?
Message Threads is currently supported on:
- Anthropic native API
- AWS (Amazon Bedrock)
- Microsoft Azure AI Foundry
- Google Cloud Vertex AI
Customers under zero-data-retention or HIPAA compliance contracts cannot use this feature at this time.
What Changes for Agent Developers?
Developers running long agent loops will feel this most directly. Applications that handle contexts in the hundreds of thousands to millions of tokens β like Claude Code itself β see the biggest benefit.
For user-facing applications, time to first token maps directly to perceived responsiveness. A 35% reduction translates to a noticeably snappier experience.
This is a beta, so testing before production deployment is advisable.
Usage Tips
- Build retry logic for 404 responses: when a thread expires mid-session, fall back to your local transcript and open a new thread.
- For short, one-shot conversations, the benefit is minimal. Reserve Message Threads for long multi-turn loops.
- For cost reduction, Prompt Caching remains the more direct lever; Message Threads addresses speed.
- Monitor thread age in long-running apps and refresh proactively if approaching the five-minute window.
Sources:
- Anthropic opens Message Threads beta, 35% faster first token β aiweekly.co
- Anthropic Opens Claude Message Threads Beta for Long Agent Sessions β tokenpost.com
- Anthropic Message Threads Beta Cuts Latency 35% on Claude β emergent.sh
- Claude API Message Threads: How the Public Beta Works β aicatchup.com
- Claude Platform release notes β platform.claude.com