Anthropic has opened Message Threads in public beta for the Claude Messages API, a change aimed squarely at the unglamorous bottleneck of long-running agent work: re-uploading an entire conversation, tool results and all, on every single turn. Send the conversation once, then only append what’s new.
How Anthropic Message Threads works
On the first request, a developer sends the full conversation (system prompt, tools, messages) with thread: {type: 'create'}. Later turns send only the new message plus the ID of Claude’s previous reply via thread: {type: 'continue', previous_message_id: ...}, and the API rebuilds the cached conversation server-side. Anthropic reports roughly 35% lower median time to first token for requests in the 500,000 to 1,000,000-token range — a figure that is Anthropic’s own, not an independent benchmark. The beta header is message-threads-2026-08-12, with support in TypeScript SDK 1.4.0 and Python SDK 2.4.0.
What doesn’t change
Pricing, rate limits and token usage are untouched — Claude still reads the full conversation each turn, and threads don’t automatically enable prompt caching. The gain is latency at response start, not total task time.
The fine print agents should read
- Threads are append-only: switching model or certain settings means starting a new thread.
- Storage is best-effort, not durable — threads expire about five minutes after the last request, so developers keep their own copy.
- Not supported with Message Batches or context editing, and not eligible for Zero Data Retention or HIPAA workloads.
Availability spans Claude Platform natively plus Claude Platform on AWS, Microsoft Foundied and Google Cloud under Anthropic management. Secondary coverage dates the launch to November 2; Anthropic’s own docs page was not directly retrievable at time of writing, so field-level details lean on technical write-ups.
