Anthropic has opened a public beta of message threads, a Messages API feature that stops long-running Claude agents from resending their entire conversation history on every turn. Announced on November 2, the feature is aimed at the sessions where latency hurts most: coding agents, computer-use agents and document-heavy workflows that stretch into hundreds of thousands of tokens.
The mechanics are simple. The first request carries the full conversation and starts a thread. Every later request sends only the new messages plus the previous response ID, using a thread parameter and a beta header. Anthropic says the full conversation stays cached on its servers between turns, so Claude can start answering sooner without waiting for a massive upload.
A 35% faster first word in very long sessions
The headline number is a median time-to-first-token improvement of about 35% on requests containing 500,000 to 1 million tokens, with message threads switched on by default for most Claude Code users. Time to first token measures the gap between sending a request and seeing the first word of a reply, and in marathon sessions that gap is dominated by re-uploading history. One analysis of Anthropic's internal testing put the effect in stark terms: the median request shrank from roughly 800 kilobytes to about 1.8 kilobytes, more than 400 times less data uploaded per turn.
The catch is that threads do not cut token bills. Claude still reads the full conversation on every turn, so standard token rates and rate limits apply unchanged. What improves is speed, not cost, and the feature is best suited to active sessions where latency matters rather than to sporadic, low-volume use.
Memory, not storage — and a five-minute clock
Threads are held in server-managed encrypted memory and are never written to durable storage, Anthropic says. Each thread expires five minutes after its last request, at which point the client must resend the full conversation to start a new one, and thread contents are deleted no more than one hour after last use. There is no API to list, read or delete threads, and applications must still keep their own copy of conversation history and a full-replay fallback for expirations and errors.
The beta also carries notable limits. Threads are not eligible for zero-data-retention or HIPAA use, do not support Message Batches, token counting, on-demand compaction or most context editing, and are unavailable to organizations with inference hooks enabled. Turns written by fallback models are not stored, so threading cannot resume from them without a fresh full-conversation request.
Available across the major clouds
Message threads is available now in public beta on the native Claude Platform, on Claude Platform on AWS, in Microsoft Foundry and on Claude Models on Google Cloud under Anthropic management, with beta support on every model. Anthropic describes the switch-on as server-side for most Claude Code users, with no minimum client version required.
The launch is Anthropic's second move in three weeks aimed at slow long sessions, following gzip compression for conversations announced on October 17. Compression shrinks the package; threads skip resending it altogether. For developers building agents that run for hours, the combination points to a clear direction: the conversation itself is becoming infrastructure, and the API is learning to remember it so the agent does not have to keep repeating itself.



