Anthropic lets Claude agents send only new messages to cut wait times

Anthropic has put message threads into public beta on the Claude Platform. The feature lets AI agents send a conversation to Claude once and then send only new messages, instead of retransmitting the full history on every request. Anthropic says that on very long sessions, with requests of 500,000 to 1 million tokens, the median wait before Claude starts replying falls by about 35%.

The change targets time, not money. Claude still reads the whole conversation on every turn, and Anthropic says prices and rate limits stay the same. For teams running long agent sessions, it is a latency tool with new state to manage. It is not a way to cut token bills.

What message threads change in the Claude API

Until now, an agent built on Claude’s Messages API has sent the whole conversation with every request. TNW notes that a coding agent can make hundreds of requests in one session, and each one normally carries every earlier message and tool result. Once that history runs to hundreds of thousands of tokens, sending it all again adds a delay before Claude starts to reply.

Anthropic’s announcement, published on 2 November 2026, describes the fix: agents send only what is new since the last turn, rather than retransmitting the full conversation and tool history each time. Anthropic has already turned threads on by default for most users of Claude Code, its coding agent.

The feature does not change which model answers, how much context it can take, or what a turn costs. It changes how the conversation reaches Anthropic’s servers between turns. In a long agent loop, that transfer can be a noticeable share of the wait.

How a Claude message thread works, request by request

According to Anthropic, developers use the feature by adding a thread parameter to Messages API requests. The flow has three stages:

1. Start. The first request includes the full conversation and starts a thread.
2. Continue. Each later request includes only the new messages and the ID of the previous response. TNW describes this as sending “only what is new, plus the ID of Claude’s last reply”.
3. Restart. Each thread expires five minutes after its last request. At that point the client must resend the full conversation to start a new thread.

Between turns, Anthropic holds threads in memory, encrypted and not written to disk. Because threads are short-lived, TNW notes that developers must keep their own copy of each conversation. The client still owns the authoritative history. The thread is a fast path layered on top of it, not a storage service.

Item	What Anthropic says
Status	Public beta
How to enable	Add the thread parameter to Messages API requests
Follow-up requests carry	New messages plus the previous response ID
Claimed latency gain	About 35% lower median time to first token on 500K–1M token requests
Token usage	Not reduced; Claude reads the full conversation every turn
Pricing and rate limits	Standard token rates and existing rate limits
Storage	In memory, encrypted, not written to disk
Expiry	Five minutes after the last request
Models	All Claude models, per TNW

The 35% latency claim and what it covers

The headline number is about time to first token (TTFT): how long a client waits between sending a request and receiving the first piece of Claude’s answer. Anthropic says median TTFT falls by about 35% on requests of 500,000 to 1 million tokens, measured with threads enabled by default for most Claude Code users. TNW reports the same figure as the “typical wait” before Claude starts answering.

Three points frame that number:

• It is Anthropic’s own measurement. Neither source reports an independent test. The figure comes from Anthropic’s rollout in Claude Code.
• It applies to very long requests. The stated band is 500,000 to 1 million tokens. Neither source gives a figure for shorter sessions. The benefit comes from not resending a large history, so smaller conversations have less to save.
• It covers the first token, not the whole reply. Generation speed after the first token is not part of the claim as reported.

For an interactive coding agent that pauses for every model call, a shorter wait before output begins can make a long session feel faster. For a batch pipeline that only cares about total throughput, the effect depends on how much of each request’s time was spent on transfer.

Speed without savings: tokens, pricing and caching

Anthropic says plainly that the feature does not reduce token usage, because Claude still reads the full conversation on every turn. Standard token rates and existing rate limits apply. The announcement did not specify pricing changes.

On cost, Anthropic’s earlier move was elsewhere. TNW reports that the company cut cache prices by 75% in September, as long agent sessions push up AI bills. Prompt caching, which stores a stable prefix of a prompt so later requests can reuse it, is the cost lever. Message threads are the latency lever. Builders running long sessions will likely weigh the two separately, and neither source describes how the two features interact.

The launch also follows a run of model releases. Anthropic’s newsroom listed Claude Opus 5.5 on 22 September 2026, Claude Sonnet 5.5 on 28 September and Claude Haiku 5.5 on 7 October. TNW describes Opus 5.5 as a cheaper model.

Memory, five-minute expiry and compliance limits

Threads add server-side state to what has been a stateless request pattern, and the design reflects that. Anthropic says threads are held in memory between turns, encrypted and not written to disk. Each one expires five minutes after its last request.

That expiry window shapes where threads fit. Anthropic says the feature is best suited to active sessions where latency matters, including long-running coding agents, computer use agents and document- or image-heavy workflows. Computer use agents are agents that operate a desktop or browser through screenshots and actions. An agent that waits longer than five minutes for a human approval, a slow tool or a scheduled job will lose its thread and need to resend everything. Client code therefore has to handle the restart path, not just the fast path.

There is also a compliance gap. TNW reports that Anthropic says threads are not eligible for its zero data retention or HIPAA health-data setups. Teams that rely on those arrangements, for example in healthcare, cannot use threads under them for now, even though thread data is not written to disk.

Where Claude message threads are available

According to Anthropic’s announcement, the feature is available now on four routes:

• the native Claude Platform
• Claude Platform on AWS
• Claude in Microsoft Foundry
• Claude Models on Google Cloud

TNW reports the beta works on all Claude models. The announcement did not give a date for general availability beyond the public beta, and did not specify model-specific limits beyond those noted.

Same-day availability on the cloud platforms matters because Anthropic beta features have not always reached every route at once. A third-party reference that tracks Anthropic’s beta flags, last checked in July 2026 and so before this launch, notes that Anthropic’s direct API may support a beta while Bedrock, Vertex, proxies and compatible gateways may not. The same reference warns that beta capabilities can change or be withdrawn, can be inconsistent across regions, models and providers, and can come with different billing and rate-limit rules. Anthropic says standard rates and limits apply in this case, but the beta label still signals that details may shift before general release.

What message threads mean for builders

For developers building long-running agents on Claude, the practical points are:

• Long sessions are where gains are claimed. The only published figure covers requests of 500,000 to 1 million tokens. Short chat turns are unlikely to see the same effect, and no data is given for them.
• Keep the full history client-side. Threads expire after five minutes of inactivity, and the client must resend the full conversation to start a new one.
• Budgets do not change. Token usage, rates and rate limits are the same. Cost reduction still depends on caching and model choice.
• Check data-handling requirements first. Threads are outside Anthropic’s zero data retention and HIPAA arrangements, per TNW.
• Measure locally. The 35% figure comes from Anthropic’s Claude Code rollout. Real gains depend on session length, tool-call patterns and how often an agent idles past the expiry window.

Open questions remain. Neither source explains how a thread behaves if a client sends a follow-up with a stale response ID, how threads interact with prompt caching, or whether the five-minute window will change after the beta. Those details matter for production agents and are likely to surface in documentation and early user reports.

What to watch next

• Independent latency tests. Third-party measurements across session lengths would show whether the about 35% figure holds outside Claude Code.
• General availability. Anthropic has not given a date. Watch for changes to the expiry window or eligibility rules at that point.
• Compliance coverage. Whether threads become eligible for zero data retention or HIPAA setups would decide adoption in regulated sectors.
• Agent frameworks and SDKs. Support in common agent tooling would make the feature easier to adopt without custom thread-handling code.