Claude Message Threads beta fixes the most annoying tax on long agent sessions: resending the entire conversation on every turn. Anthropic launched message threads in public beta on November 2, 2026, for the Claude Messages API, and the headline number is a median time-to-first-token drop of about 35% on requests containing 500,000 to 1 million tokens.
That number only applies to very long sessions. But if you run coding agents, computer-use agents, or document-heavy workflows, those are exactly the sessions that feel slow — and the fix is one new parameter.
How message threads work
Until now, an agent sent the full conversation history, including tool results, on every turn. With message threads, the first request includes the full conversation and starts a thread. Each later request includes only the new messages and the ID of the previous response.
Anthropic keeps the thread state server-side, in encrypted memory, and rebuilds the full context for Claude on each turn. In long Claude Code conversations, Anthropic reports the median request shrinking from about 800 KB to 1.8 KB. Smaller uploads mean less serialization and upload time before Claude starts generating.
Three limits define how to use it safely:
- Best-effort, not durable. Threads expire five minutes after their last request, and there is no API to list, read, or delete them.
- Keep your own copy. If the thread is gone, your client resends the full conversation to start a new one. Your conversation still needs a home of its own.
- No bill cut. Claude still reads the full conversation on every turn, so token usage, standard token rates, and existing rate limits are unchanged.
The evidence: what Anthropic actually announced
| Detail | What the source says |
|---|---|
| Launch | Public beta, November 2, 2026, on the Claude Messages API |
| Latency claim | Median time to first token about 35% lower on 500,000–1,000,000 token requests |
| Payload example | Median Claude Code request from about 800 KB to 1.8 KB in long conversations |
| How to enable | Add the thread parameter; beta header message-threads-2026-08-12; SDKs provide a helper |
| Already on | On by default for most Claude Code users, per Anthropic |
| Thread expiry | Five minutes after the last request; contents deleted no more than one hour after last use |
| Storage | In memory under server-managed encryption, never written to durable storage |
| Pricing | Standard Claude token rates and existing rate limits apply; no change |
| Where available | Native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, Claude Models on Google Cloud (Anthropic managed) |
| Not supported | Message Batches, token counting, on-demand compaction, most context editing, Inference hooks orgs |
| Not eligible | Zero data retention and HIPAA use |
Time to first token (TTFT) is the wait before a model starts producing its first word. For an agent loop, that wait repeats on every turn, so cutting it compounds across a long session.
Adopt message threads without losing a session
1. Store every message locally — Your log is the recovery path, not the thread.
2. Send full, then deltas — First request starts the thread; later requests send only what’s new.
3. Handle the 404 — On expiry, resend the full conversation and start a new thread.
4. Measure TTFT, not tokens — Your bill will not fall; your wait should.
5. Skip regulated workloads — No zero-data-retention or HIPAA use for threads.
How to profit from it
Speed is a product feature when an agent sits in front of a paying customer or a developer waiting on a build. Here is how to turn a latency cut into revenue:
- Sell faster agent sessions. If you run a coding or document agent for clients, threads cut the dead wait on long jobs. Measure before and after, and show the client the TTFT graph.
- Offer a long-session audit. Find clients whose agents resend huge histories every turn, enable threads, and charge for the optimization — not for a token saving that does not exist.
- Build for active sessions only. Threads suit always-on coding and computer-use agents. A chatbot that sleeps for hours will expire its thread anyway, so do not promise speed there.
- Do not quote a price cut. The bill stays the same. Anyone selling threads as a cost saving is misreading Anthropic’s announcement.
- Charge for the fallback. The valuable engineering is the local log and the resend-on-expiry path. That reliability work is billable, and it is what keeps a five-minute expiry from becoming an outage.
Sources
Anthropic’s official announcement, Make long-running agents faster with message threads (November 2, 2026), plus coverage by Unite.AI and DEV. The 35% figure is Anthropic’s median for very long requests, not a guarantee for short chats.
Frequently Asked Questions
What are Claude Message Threads?
A public-beta Messages API feature that keeps active conversation state server-side, so clients send only new messages after the first full request.
How much faster are agents with threads?
Anthropic reports median time to first token about 35% lower on requests with 500,000 to 1 million tokens, based on Claude Code use.
Do message threads lower my Claude bill?
No. Claude still reads the full conversation each turn, so token usage, rates, and rate limits are unchanged.
How long does a thread last?
It expires five minutes after its last request, and contents are deleted no more than one hour after last use.
How do I enable message threads?
Add the thread parameter to Messages API requests with the message-threads-2026-08-12 beta header; official SDKs provide a helper.
Where is the beta available?
On the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud under Anthropic management.
Can I use threads for HIPAA or zero-data-retention workloads?
No. Threads are not eligible for zero data retention or HIPAA use.
Keep Reading
Want more tools that actually pay for themselves? Browse the AI Tools section for the latest launches, pricing breakdowns, and profit playbooks — and check the official source linked above before you buy or deploy anything.