Claude Code Starts Answering 35% Sooner on Long Jobs, Anthropic Says

Anthropic says Claude Code's biggest requests now start answering about 35% sooner, because Claude Code stops re-sending the whole chat on every turn. On November 2, 2026, the company put message threads into public beta on the Claude Platform. With message threads on by default for most Claude Code users, Anthropic reports that requests holding 500,000 to 1 million tokens (a token is a small chunk of text, about three quarters of a word) saw median time to first token, the wait before the first word of the answer appears, drop by about 35%.

Until now, an agent (a program that works through a task in many back-and-forth steps) uploaded every earlier message each time it asked Claude for the next step, sometimes hundreds of times a session. With a thread, the first request sends the full conversation, and later requests send only the new messages plus the ID of Claude's last reply. Claude still reads the whole conversation every turn, so token use stays the same and standard token rates and rate limits apply. What shrinks is the upload and the checking work the API does before the model starts.

Jon's agents run in Claude Code around the clock on always-on Mac Minis, in long sessions with big contexts, so this touches his work directly. The fleet runs Claude Code 2.1.318, whose code carries the threads beta name. It has not timed threads itself, so the 35% is Anthropic's number.

On November 3, a ZDNET analysis republished on Yahoo by David Gewirtz went back to about 980 turns in his Claude Code logs from before the feature. He saw waits of about 3 seconds on contexts of 500,000 tokens and up, and 100 to 225 model calls in four hours. Applying Anthropic's 35% gave an estimate of 2 to 4 minutes saved per four-hour session: math on old logs, not an on-and-off test.

How to use it

Claude Code users do not need to do anything: Anthropic says the feature is already on for most of them. The Claude Code environment variables page lists no switch for threads.

For a developer running an agent loop on the API, the message threads documentation gives the steps:

1. Send the beta header anthropic-beta: message-threads-2026-08-12 on every request.
2. On the first turn, send the full conversation with system and tools, plus "thread": {"type": "create"}.
3. On each later turn, send only the new messages, leave out system and tools, and add "thread": {"type": "continue", "previous_message_id": "<last response id>"}.
4. Or let the SDK do it: client.beta.messages.thread(...), then thread.send(...) each turn.

The catches: a thread is best-effort and expires 5 minutes after its last request, and a 404 error means re-sending everything. It is append-only, each reply ID can be continued once, and fields like model cannot change. No response confirms a turn was stored, and the SDK keeps its copy only in memory, so an app needs its own saved history. Prompt caching is still set up separately. There is no API to list, read or delete a thread. Threads are not eligible for zero data retention (Anthropic keeping nothing after a request); the docs say contents stay encrypted in memory and are deleted within an hour. Organizations with HIPAA readiness or inference hooks turned on cannot use them.

What nobody knows yet

The 35% is Anthropic's own median, on its own traffic, for very large requests only. Nobody has published a clean on-and-off test. Anthropic says "most" Claude Code users have it, without saying who does not or naming a version. Small conversations should gain little. There is no date for the end of the beta.

What to watch

The next signs are a Claude Code changelog entry naming threads, an independent on-and-off timing test, and a newer beta header or a general release, which would mean the request format is settled. No date for any of them has been announced.