Anthropic: Message threads promise faster first responses in very long AI sessions

What's new

Anthropic has made message threads available in public beta on Claude Platform. The launch is dated November 2, 2026. The feature is aimed at agents, meaning AI programs that work autonomously for long stretches and take many steps in a row.

Previously, an agent like this had to send the entire conversation history again on every turn. With message threads, the first request contains the whole conversation, while the following ones contain only the new messages and the ID of the previous response. Developers must opt in to the feature via the thread parameter in the Messages API, while according to Anthropic it is turned on by default for most Claude Code users. According to Unite.AI, the feature applies to all Claude models during the beta period.

What makes it smart

Think of it as no longer having to hand over your entire folder of papers every time you ask a new question. Now you hand over only what's new and point to where you left off.

This makes a particular difference when there is a lot to send: large documents, images or screenshots. Anthropic highlights three types of use: long-running coding agents, computer use agents (which control a computer and no longer have to upload screenshots over and over) and workflows involving many documents or images.

Message threads are turned on by default for most Claude Code users. The sources do not say whether this applies to anyone else.

Cheaper or better?

According to Anthropic's own measurements, the median time to first token (that is, how long it takes before the response starts to arrive) drops by about 35 percent for requests of 500,000 to 1 million tokens. The figure relates to Claude Code and applies only to very long sessions. There is no independent verification.

It does not get cheaper in terms of tokens: token consumption is unchanged, because Claude still reads the entire conversation on every turn. Only the request itself gets smaller. On the other hand, standard token prices and existing rate limits apply, and no extra fee for message threads is mentioned. The sources do not compare competitors.

There are also limitations. Threads do not cache prompts automatically, so developers must specify cache breakpoints themselves, with a maximum of four per request. Message Batches, token counting and most forms of context editing are not supported, and the feature is not available to organizations with Inference hooks turned on.

Who it's good for

The feature is best suited to developers and companies building agents with long sessions, such as coding assistants, agents that operate a computer, or solutions that work with many documents and images.

You do need to know the rules for data, though. Between turns, threads are held in memory, encrypted and never written to disk. They expire automatically five minutes after the most recent request and are deleted after one hour at the latest. Threads are not covered by zero data retention or HIPAA compliance, which may rule out the feature for regulated users. The client must keep its own copy of the messages. Threads are append-only, and each response ID can be continued only once.

Message threads are in beta on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude models on Google Cloud. The sources do not specifically mention availability in Denmark or the EU.