Skip to content
NewsProduct launch

Anthropic lets Claude agents skip resending chat history, cutting latency up to 35%

· by

The short version

Anthropic's new message threads beta holds a Claude API conversation in server memory between turns, dropping median time-to-first-token by about 35% on requests between 500,000 and 1 million tokens.

Anthropic lets Claude agents skip resending chat history, cutting latency up to 35%

Long-running coding agents on the Claude API no longer have to resend their entire conversation on every single turn. Anthropic's new message threads feature, in public beta since November 2, 2026, holds a session in server memory between requests and cuts median time-to-first-token by about 35% on requests between 500,000 and 1 million tokens, per Anthropic's product announcement.

What happened

Message threads add a thread parameter to Messages API requests. The first request in a session includes the full conversation, system prompt, and tools, and starts a thread; every request after that sends only the messages added since the last turn plus the ID of the previous response, Anthropic's documentation explains. The API rebuilds the full conversation on the server before it calls the model, so token usage and billing stay the same even though each request is smaller. A thread lives in memory, encrypted and never written to disk, and expires five minutes after its last request; after that, a client has to send the full conversation again to start a fresh one, per Anthropic. The feature ships behind the message-threads-2026-08-12 beta header and is already on by default for most Claude Code users, according to the documentation.

HIPAA and zero-retention customers cannot turn it on yet

Message threads work on the direct Claude API, Claude Platform on AWS, Claude Models on Google Cloud, and Microsoft Foundry, but Anthropic's own feature table marks the beta not eligible for zero-data-retention agreements and not HIPAA eligible, per the documentation. That excludes exactly the regulated enterprise and healthcare deployments most likely to run the hundreds-of-thousands-of-token agent sessions the feature targets. Anthropic also calls the thread best-effort: a continue request returns a 404 if the server-side copy expired or got evicted, so an application still has to keep its own copy of every message and be ready to replay the whole conversation into a new thread. Teams building long coding or computer-use agents get a latency win from one added parameter, but they still own the conversation state and the fallback path if a thread drops mid-session.

What to watch next

Watch whether Anthropic extends zero-data-retention and HIPAA eligibility to message threads as the beta matures, and whether the message-threads-2026-08-12 header gives way to a stable, unversioned parameter once the feature reaches general availability.

Sources