Claude message threads: about 35% faster on 500k-1M tokens

Anthropic on November 2 opened message threads in public beta on the Messages API. With it, an agent sends only what is new since its last turn instead of the whole conversation, and the feature is already on by default for most Claude Code users. If your agents run long sessions, you get a shorter wait for the first token and the same bill. In Claude Code, median time to first token on requests of 500k to 1M tokens fell about 35%, but Claude still reads the full conversation every turn, so token usage stays the same.

How to turn it on

Most Claude Code users don’t need to do anything. In your own agent, add the thread parameter to Messages API requests. The first request carries the full conversation and starts a thread. The News Stack adds that this first request includes the system prompt and tool definitions. After that, each request sends only the new messages and the ID of the previous response. Standard Claude token rates and existing rate limits apply.

Your client does have to handle expiry. A thread lapses 5 minutes after its last request, and the client then resends the full conversation to start a new one. Between turns, Anthropic keeps each thread in memory, encrypted and never written to disk. The Claude Platform docs add that a thread is deleted no more than 1 hour after its last use.

Why the wait shrinks and the bill doesn’t

The Messages API is stateless: without threads, every request carries the entire history again. A coding agent can make hundreds of requests in one session, and once the conversation reaches hundreds of thousands of tokens, resending it adds time before Claude starts answering. The News Stack describes where that time goes. The API has to validate and tokenize all the incoming data before Claude can respond. With a thread, Anthropic rebuilds the full conversation on its own servers before passing it to the model.

So the speedup comes only from transfer and preprocessing, and the model’s side of the work doesn’t change. Anthropic says so directly: “While each request is smaller with message threads, Claude still reads the full conversation on every turn, so token usage stays the same.”

ZDNET‘s author timed about 980 turns in their own Claude Code logs from before the feature. At contexts of 500,000 tokens and up, Claude usually began a response about 3 seconds after each mid-task request. Claude applied Anthropic’s 35% median to those logs and estimated a saving of 2 to 4 minutes per four-hour work session. The 35% itself is Anthropic’s measurement.

Prompt caching is still what cuts cost. It has been around for about two years and works alongside threads, the same ZDNET piece notes. In August, Anthropic’s engineering blog put cache reads at 0.1x the input price. It also said the cache expires after five minutes on an API key, the same window a thread lives after its last request. Going by those August figures, an API-key session that sits idle for more than five minutes loses both the cache and the thread.

Where it pays off

Anthropic says the more context an agent carries, the more threads can cut time to first token. It names long-running coding agents, where every tool result or file edit adds to the conversation, and document- and image-heavy workflows, where requests get smaller. For computer use agents, screenshots stay with the thread instead of being re-uploaded on every turn. Outside Claude Code, message threads is available now in public beta on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed).