Claude caches long conversations on the server, cutting requests from 800 KB to 1.8 KB

Claude's Messages API has added message threads, in public beta.

Previously, every request had to resend the entire conversation history. Now the API caches the conversation on the server, and the client only sends the new message plus the previous message ID, letting the model continue from the existing context.

The official numbers are very specific. Claude Code has already integrated message threads. For requests of 500K to 1M tokens, median time to first token is about 35% lower. In long conversations, the median request body drops from about 800 KB to 1.8 KB.

Usage is not complicated: add the thread parameter and beta header to the request. It is available in Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude on Google Cloud, billed at standard token prices. Full details are in this blog post.

But don't treat it as permanent memory. Some commenters noted in replies that threads expire after 5 minutes of inactivity, and the client still has to save history itself. It solves the problem of repeatedly uploading the conversation every time, not long-term memory hosting on your behalf.

Another easily confused point is prompt caching. Caching has long reduced the cost of repeatedly sending long histories. The more direct change from threads is a smaller request body and faster time to first token. Some commenters summarized that prompt caching saves compute cost, while threads save request size and latency. They are not the same thing.

A developer added that the TypeScript SDK's toolRunner already runs by thread on supported clients, sending only new messages each turn; passing thread: false falls back to the old full-history behavior.

Some joked that over the past two years many people wrote SQLite wrappers to manage context arrays; now Anthropic has put the state on the server, so clients carry one less burden. As for whether that burden will automatically disappear after 5 minutes, it depends on how you design retries and persistence.