# Claude's message threads make long agents faster, not cheaper

URL: https://www.thedeepfeed.ai/posts/2026-11-05-claude-message-threads-latency-not-cost/
Category: Tools
Published: 2026-11-05
Author: the-deep-feed
Tags: anthropic, claude, api, agents, latency
Kind: news

> Anthropic's new Messages API beta stops agents re-uploading the whole conversation. It cuts time to first token on huge contexts, but billed tokens and rate limits do not change.

## TL;DR

- Anthropic's **message threads** beta (Nov 2) lets an agent send only new messages plus a reference to the previous response.
- For Claude Code requests of **500k to 1M tokens**, Anthropic says median time to first token fell **about 35%**.
- Claude still reads the full conversation each turn, so **token usage, pricing and rate limits stay the same**.
- Threads expire **5 minutes** after the last request; clients must keep history to replay on a 404.
- Not listed on Amazon Bedrock, and the docs mark it **not ZDR or HIPAA eligible**.

Anthropic's new message threads beta fixes a transport problem, not a cost problem. Long agent loops no longer have to re-upload hundreds of thousands of tokens on every call, which makes them start answering sooner. The bill for those tokens is unchanged, and teams with zero-data-retention or HIPAA requirements can't use it.

# What shipped

On November 2, Anthropic added [message threads](https://platform.claude.com/docs/en/release-notes/overview) to the Messages API behind the `message-threads-2026-08-12` beta header. The [docs](https://platform.claude.com/docs/en/build-with-claude/message-threads) describe the pattern: the first request carries the full conversation, and each later request sends only the new messages and the ID of the previous response. The server rebuilds the full context before Claude sees it.

According to [Anthropic's announcement](https://claude.com/resources/articles/message-threads), the feature is already on for most Claude Code users: *"With message threads on by default for most Claude Code users, requests with 500k to 1M tokens see median TTFT drop by about 35%."*

# Why the bill doesn't move

The saving is upload and request validation time, not model compute. Anthropic says so plainly, as [The New Stack](https://thenewstack.io/anthropic-claude-message-threads/) also points out: "While each request is smaller with message threads, Claude still reads the full conversation on every turn, so token usage stays the same." Standard rates and rate limits apply.

Threads also don't cache the prompt for you. The [docs](https://platform.claude.com/docs/en/build-with-claude/message-threads) say teams that want cheaper repeated context still have to set `cache_control` separately. Threads cut latency, and prompt caching is still what cuts cost.

# The operational fine print

- **Short lifetime.** A thread expires five minutes after its last request and is deleted no more than an hour after its last use. An expired thread returns a 404, and the client has to replay the full history. In practice, agents still need to store the whole conversation locally.
- **Append-only.** Each response ID can be continued only once, so branching a conversation means starting again from the full history.
- **Where it runs.** The docs list the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry. Amazon Bedrock is not listed.
- **Compliance.** The docs mark the feature as not ZDR eligible and not HIPAA eligible.

# Our read

This is the clearest sign yet that the bottleneck for long-running coding and computer-use agents is moving from model speed to data plumbing. The New Stack cites an estimate by StackOne CTO Guillaume Lebedel: in a 50-step screenshot-heavy session, cumulative screenshot uploads would fall from about 191 MB to 7.5 MB. That figure comes from a simulation, not a live test. If you run long agent loops on the first-party API, turn it on. If your cost model assumed "less data sent" means "fewer tokens billed", it doesn't. For the bigger picture on coding-agent platforms, see [coding agents pick sides](/posts/2026-10-16-coding-agents-pick-sides/).

## Sources

- [Anthropic — Make long-running agents faster with message threads (Nov 2, 2026)](https://claude.com/resources/articles/message-threads)
- [Claude Docs — Message threads](https://platform.claude.com/docs/en/build-with-claude/message-threads)
- [Claude Platform — Release notes (Nov 2, 2026)](https://platform.claude.com/docs/en/release-notes/overview)
- [The New Stack — Anthropic speeds up Claude Code with message threads. But your token bill stays the same. (Nov 3, 2026)](https://thenewstack.io/anthropic-claude-message-threads/)
- [DEV Community — Claude message threads: send only new messages on long agent loops (Nov 3, 2026)](https://dev.to/sueun-dev/claude-message-threads-send-only-new-messages-on-long-agent-loops-4b0g)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-11-05-claude-message-threads-latency-not-cost/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt