Anthropic cuts Claude agent latency 35% with message threads

If you have ever watched an autonomous coding agent work through a massive project, you know the waiting game. Dozens of tool calls in, the agent starts crawling to a halt before spitting out a single character.

Anthropic wants to fix that drag. The company just rolled out Message Threads in public beta for its Claude Messages API, aiming to make long-running agent loops significantly snappier.

Basically, your client no longer has to dump the entire chat history back across the wire on every single turn.

Stop re-uploading the whole conversation

A typical coding session can fire hundreds of requests back and forth. Every terminal log, file edit, and tool result keeps piling up. Once a session swells past hundreds of thousands of tokens, just transmitting that massive payload creates a noticeable delay before Claude can even begin to think.

With Message Threads, you send the full context once to initiate a thread. From that point on, your client only passes the thread parameter, the previous response ID, and whatever new messages just happened.

Anthropic says Message Threads is already on by default for most Claude Code users. For hefty sessions sitting between 500k and 1M tokens, the median time to first token (TTFT) drops by roughly 35%.

Under the hood, Anthropic holds each active thread encrypted in memory without touching disk storage. If five minutes tick by without a new request, the thread simply expires, meaning your client will just re-upload the full history on the next turn to start a fresh thread.

Developers can test this right now across the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Google Cloud managed instances. Keep in mind that token usage remains identical and standard rates still apply, because Claude still reads the full context on each pass.

Why it matters

We think this is a textbook example of unglamorous infrastructure work that makes a massive difference in day-to-day usability.

If you build autonomous agents or lean heavily on Claude Code, shaving more than a third off initial latency fundamentally changes how interactive an agent feels. It is especially huge for computer use agents that constantly juggle heavy screenshots, where uploading high-resolution visual history over and over was simply wasteful.

The five-minute in-memory expiration is a clever design choice. It dodges the security and regulatory headaches of persistent conversation storage while keeping active agent loops blisteringly quick.

The catch, of course, is your bill. Because the model still processes every token under the hood, this will not save you a penny on API costs. But as context windows stretch into millions of tokens, the teams that solve data plumbing and latency bottlenecks are the ones who will actually make complex agents practical.