Claude's new message threads cut time to first token by about 35% on 500k to 1M token requests. Token usage stays the same

A faster agent loop with one new parameter, as long as you build for the thread disappearing.

What happened

Anthropic released message threads in public beta for the Messages API. Normally every request resends the whole conversation. With a thread, the first request sends the full conversation and creates the thread, and each later request sends only the new messages plus the ID of the previous response.

Anthropic says that with threads on by default for most Claude Code users, requests with 500k to 1M tokens saw median time to first token drop by about 35%. Token usage does not change, because Claude still reads the full conversation every turn. Standard token rates and rate limits apply. Threads are held in memory, encrypted and not written to disk, and expire 5 minutes after the last request. It works on the Claude Platform, Claude Platform on AWS, Microsoft Foundry and Google Cloud.

The docs add the details that matter in production. A thread is best effort: when it cannot be continued, a continue request returns a 404 and you resend the full conversation in a new create request. Your application stays the source of truth, so keep your own copy of every message. System prompt and tools go only on the create request. A response written by a fallback model is not stored, so you cannot continue from it. Each request needs the message-threads-2026-08-12 beta header.

My take

This is for long agent loops where latency hurts: agents that read a lot of documents, screenshots or tool results across many turns. For a short three step workflow, it changes little.

Two points to be clear on with clients. First, it is a speed feature, not a cost feature. Your token bill stays the same. Second, the 5 minute expiry means any loop that waits on a human, a webhook or a slow API will lose its thread.

How I would add it to an existing agent:

1. Keep storing the full message history in your own database, as you do today.
2. Wrap the call so a thread_not_found 404 triggers a full replay automatically.
3. Do not change system prompts or tools mid run. Start a new thread instead.
4. Log time to first token before and after, so you know it helped on your workload.