Anthropic cuts Claude's time to first token by 35 percent for long sessions

A coding agent sends the whole conversation to the model again every turn. After an hour, that turn can carry hundreds of thousands of tokens the server has already read. Anthropic has released a feature that stops the resend. The company says long Claude Code sessions now start replying about 35 percent sooner.

The feature is called message threads. It went to public beta on 2 November. It lives on the Messages API, which powers Claude Code and third-party Claude agents. Anthropic says it is already on by default for most Claude Code users.

What message threads changes

Before the feature, every request carried the full chat. That meant every message, every tool result and every file the agent had read. On a long coding session, each request grew to hundreds of thousands of tokens. The API still had to validate and tokenize all of them before the model saw the first new word. Anthropic says the wait between send and reply grew with the chat.

With message threads, the first request creates a thread. It still carries the full chat, the system prompt and any tools. Each later request carries only the new messages and the id of the last reply. The server keeps the chat in memory and rebuilds it before each call to Claude.

Anthropic says the gain is in time to first token. That is the delay before the model starts writing. For Claude Code requests between 500,000 and one million tokens, Anthropic says the median delay drops by about 35 percent. The model still reads every turn. Token counts and the bill do not fall.

What is kept and what is not

A thread lives for five minutes after the last request. The server deletes it no later than one hour after the last use. The API stores the chat encrypted in memory and writes nothing to disk. If a continue request arrives after the thread is gone, the server returns a 404. The client then sends the full chat again in a new create request. Anthropic says the SDK thread helper handles this replay. But the client must keep its own copy of every message. No response says whether the API stored that turn.

The feature works with prompt caching. Anthropic says cache breakpoints still apply. A team that wants lower cost as well as lower latency can turn both on.

Where the gain shows first

Anthropic names three cases it says will benefit most. Long coding agents that collect tool results and file edits. Computer use agents that repeat screenshots every step. Workflows that send many documents or images.

Guillaume Lebedel, cofounder and CTO of StackOne, ran a simulation and shared the numbers with The New Stack. He assumed one 150 KB screenshot per step. Across fifty turns, the agent would upload about 191 MB of screenshots without threads. With threads, that figure drops to about 7.5 MB. Lebedel said the numbers are from a simulation, not a live run. He expects computer use agents to adopt the feature first, with most having it on by end of December.

Message threads runs on the native Claude Platform, AWS, Microsoft Foundry and Google Cloud. The usual Claude token rates apply. The feature is a beta.

For a team building long-running agents, the move is to turn the feature on. Pair it with prompt caching when cost matters. Keep a local copy of every message so a dropped thread never loses state. The upload was the slow part. The bill is still the token count.

Source

Anthropic on the public beta, and The New Stack by Paul Sawers with the StackOne simulation figures.