Anthropic has a beta feature for Claude agents that the company says delivers faster time to first token

What's new

Anthropic has a new beta feature for developers called Message Threads. It is meant to give long-running agents, meaning AI programs that work independently across many steps and turns, a faster time to first token. That is the time between sending a message and the first piece of the reply appearing. The improvement is meant to be most noticeable in long conversations.

The feature is enabled with the beta header message-threads-2026-08-12. It is listed as available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic-operated) and Microsoft Foundry. The sources do not say whether it is open to users in Denmark and the EU on special terms.

What's smart

Without a thread, the program resends the entire conversation on every new turn. The API then has to validate and tokenize everything again. Tokenization is when the text is split into the small pieces the model works on. The longer the conversation, the more work there is in repeating that every time.

With a thread, only the new messages are sent. Most of the work therefore concerns only the new turn. According to Anthropic, requests get smaller, and time to first token stays lower and more stable as the conversation grows. That is Anthropic's own description. In the portion available, the documentation does not give an independently measured figure.

The thread's contents, meaning the conversation, system prompt and tools, are held in memory, encrypted with a server key that Anthropic controls. They are never written to persistent storage and are deleted no later than 1 hour after the thread's last use.

Cheaper or better?

Pricing has not been disclosed. The sources do not say whether the feature costs extra or whether special usage limits apply. There is also no comparison with similar features from competitors.

What the feature promises is better speed rather than a lower price: less data per request and a response that does not get slower and slower the longer you talk.

Note that the feature is not eligible for Zero Data Retention (ZDR), an arrangement in which data is not stored at all after use. Companies with strict data-handling requirements should factor that into their decision. Using threads does not affect other features that are ZDR-eligible.

What it is good for

The feature is for developers and companies building AI agents that work for a long time and go many rounds with the model. That could be, for example, an assistant that solves a task in many steps, or a customer service agent with a long conversation history.

It supports several generations of Claude models, including Opus 5.5, Sonnet 5.5 and Haiku 5.5, as well as older models such as Sonnet 4.5 and Haiku 4.5.

For ordinary users of Claude as a chat app, this is a behind-the-scenes feature. Anthropic has not said whether ordinary users will notice a difference.