Claude gets message threads: Agents no longer have to resend everything

What's new

On November 2, 2026, Anthropic launched a feature called message threads in public beta. Until now, an application talking to Claude has had to upload the entire conversation history on every single turn. With threads, the application first sends the whole conversation and then only the new messages.

Technically, a thread is temporary storage of the conversation at Anthropic. The developer sends the new messages and an ID from the previous response, so Claude knows where the conversation left off. In Claude Code, threads are turned on by default for most users.

The feature is available on Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude models on Google Cloud that are managed by Anthropic. Beta support applies to all models.

What's smart about it

Think of an agent that works for hours: a coding agent digging around in a large project, or a computer-use agent that controls a screen and takes screenshot after screenshot. Without threads, the whole pile of text and images has to be sent back and forth again and again. With threads, the screenshots stay in the thread, and only the new material has to travel over the network.

Anthropic says the median time to first token drops by about 35 percent for Claude Code requests of 500,000 to 1 million tokens (chunks of text) when threads are turned on. That is Anthropic's own measurement and has not been independently confirmed. It applies only to the very large requests.

Cheaper or better?

Faster for very large Claude Code requests, according to Anthropic's own measurement, but not cheaper. There is no separate price for message threads, and standard token prices and rate limits apply. Token usage, and therefore the bill, does not change, because Claude still reads the entire conversation on every turn. Specific prices are not given in the sources, and there is no comparison with competitors.

There are also limitations. Storage is best effort, so developers must save the conversation history themselves and be able to replay it. Threads are append-only: if you switch models or settings, you have to start a new thread. Threads expire five minutes after the most recent request, and according to the documentation, content is deleted no later than one hour after last use.

The feature does not meet HIPAA requirements and has no zero data retention, meaning no guarantee that data is not stored. It also cannot be used with Message Batches or in organizations with inference hooks. The sources do not say whether the feature is available in Denmark or the EU.

Who it's good for

Message threads are primarily for developers and companies building long-running agents. The sources point to coding agents, computer-use agents and workflows with many documents and images.

For ordinary users of Claude in a chat, there is nothing to do themselves. The gain comes if the tools they use are built on the feature and therefore respond faster during long workflows. Companies with strict data protection requirements should note that the feature is neither HIPAA-compliant nor free of data retention.