Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Message Threads cut Claude Code waits by a third

Written November 6, 2026

Anthropic's server-side message threads hold conversation history between turns. On long sessions the first token arrives about 35% sooner.

What Message Threads is

The Messages API has always been stateless: every turn resends the entire conversation, and the API validates and tokenizes all of it again before Claude starts answering. Message threads, in public beta since November 2, 2026, lets the API keep the conversation on the server between turns. The first request sends the full conversation, the system prompt, and the tools, and starts a thread. Each request after that sends only the new messages and the ID of the previous response, and the API rebuilds the conversation before it calls the model. Between turns, the thread is held in memory, encrypted, and never written to disk.

The point is latency, not tokens. Sending less per turn means less validation and tokenization before the first token, and the wait before Claude starts answering shrinks — most on the longest sessions, where the resend was the slowest part of the request.

What it does for Claude Code

Message threads is already on by default for most Claude Code users, server-side. Nothing needs updating and no minimum client version is named: the feature runs in Anthropic's infrastructure, not in the installed client. On requests of 500,000 to 1 million tokens, the median time to first token falls by about 35%. That figure is Anthropic's own, measured on its Claude Code rollout — not an independent benchmark, and not a figure for shorter sessions. Anthropic gives no number below 500,000 tokens, saying only that the benefit grows with the amount of context an agent carries.

The practical effect is narrower than the headline sounds. The 35% is a cut in the wait before the first token appears, on very long requests. It is not a claim that every Claude Code request gets 35% faster, and it is not a claim about total response time — generation speed after the first token is unchanged.

What it does not do

It does not cut the bill. Claude still reads the full conversation on every turn, so token usage stays the same, and standard token rates and existing rate limits apply. The documentation states it plainly: threads do not change pricing or rate limits. A thread does not cache the prompt for you either — prompt caching still needs to be requested separately.

It is also not a durable store. A thread expires 5 minutes after its last request, and the API deletes it no more than 1 hour after its last use. Threads are best-effort: a continue request against a thread that is gone returns a 404, and the client resends the full conversation to start a new one. Each response ID can be continued exactly once, and a thread is append-only — to replace earlier turns with a summary, start a new thread from the summary.

What API builders need

For direct use of the Messages API, the requirements are small but strict. Every request of a thread must carry the message-threads-2026-08-12 beta header, and each request adds a thread parameter: {"type": "create"} on the first turn with the full conversation, then {"type": "continue", "previous_message_id": "<previous id>"} with only the new messages. System and tools go on the create request only — sending either on a continue request returns a 400 error. The SDKs include a thread helper that builds these requests, adds the beta header, and replays the full conversation on a 404.

Three constraints decide whether threads fit an application. Your application stays the source of truth for the conversation: keep your own copy of every message and build the replay path first, because no response says whether the API stored a turn — you find out on the next continue request. Threads are not eligible for zero data retention, and organizations with HIPAA readiness enabled cannot use them. And threads do not work with Message Batches, token counting, on-demand compaction, or most context editing. Threads are in beta on every model, on the native Claude Platform and on Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed).

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Claude Code subagent updates now cost 99% less · Baseline usage before the shared allowance · Run parallel subagents in Claude Code