Anthropic's API can now keep a conversation on its servers between turns, and Claude Code already uses it to start long replies about 35% faster.

Anthropic added message threads to the Claude Messages API as a public beta on Monday (ClaudeDevs on X). Normally an app has to send the whole conversation back with every new message. With a thread, the API keeps the conversation on its side, so you send only the new message plus the ID of the last one, and Claude "starts responding sooner." Anthropic says it has already built threads into Claude Code. On requests of 500K to 1M tokens, the median time before the first words appear dropped by about 35%, and the median request in long conversations shrank from about 800 KB to 1.8 KB (ClaudeDevs reply on X). Threads are on by default for most Claude Code users (Anthropic article). It's a speed change, not a discount. Claude still reads the full conversation every turn, so token use stays the same, and the docs say "threads do not change pricing or rate limits" (Claude docs). For developers, it's turned on with the beta header message-threads-2026-08-12. A thread expires 5 minutes after its last request. If it's gone, the API returns a thread_not_found error and you resend the full conversation; Anthropic's SDK helper handles that for you. Threads don't work with the Batches API, token counting or compaction, and they aren't covered by zero data retention or HIPAA. They're available on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

When you'll feel it: if you use Claude Code, you may already be feeling it. Long sessions should start answering a bit sooner, with no setting to change. Your usage limits and token counts won't move. If you build on the API, it's worth testing on long chats. Keep your full-history code path, since threads expire after 5 minutes of quiet.