Anthropic launches message threads in public beta to cut agent wait times in long sessions
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Anthropic has started a public beta of message threads on Claude Platform: an agent sends the whole conversation once and then only what is new each turn, aimed at very long sessions. Anthropic estimates that for requests of 500,000 to 1M tokens, the median wait before the first reply falls about 35%. Token usage and rates stay as they were.
Anthropic has opened message threads in public beta for the Messages API on Claude Platform, its developer platform. Rather than resending the whole conversation with every request, an agent sends the full history once and from then on only the new messages, along with the ID of the previous response.
The problem it targets is long sessions. A coding agent can make hundreds of requests in one session, and each one drags along every earlier message and tool result; once a conversation reaches hundreds of thousands of tokens, resending it delays the moment Claude starts to respond. According to Anthropic, with the feature on by default for most Claude Code (its coding tool) users, requests of 500,000 to 1 million tokens take about 35% less time, at the median, to deliver their first token.
To use it, developers add a thread parameter to their requests. Anthropic holds each thread in memory, encrypted and never written to disk, and lets it expire 5 minutes after the last request; the client then has to resend the entire conversation to start a new one. Even though each request weighs less, the model still receives the complete conversation on every turn; that is why token consumption is the same and the usual rates and limits apply. The company recommends it for coding agents that work for a long time, for those that operate a computer (screenshots stay in the thread instead of being uploaded again) and for tasks loaded with documents or images.
The beta is available on Claude Platform and on its versions for AWS, Microsoft Foundry and Google Cloud (the latter managed by Anthropic). According to Unite.AI, which goes through the technical documentation, each application has to store the message history on its own, threads fall outside zero data retention and HIPAA use, and only one continuation can come out of each response.
Why it matters · analysis and opinion
The change does not make Claude think less: it spares the client from packaging and resending what it already said. It is plumbing, but in agents that chain hundreds of calls, that plumbing sets how snappy the tool feels to whoever uses it. There are trade-offs to weigh before adopting it: it is still in beta, threads expire quickly and the application must keep its own copy of the conversation to reopen them, so the complexity does not vanish, it moves into the client code. Nor does it make anything cheaper, because token consumption is the same, and anyone handling sensitive data should check that the threads' retention terms fit their requirements. A team already building coding agents can try it by measuring with its own long sessions.
Official source: Anthropic · Written with the help of AI: how we make the news
