Anthropic has introduced message threads in public beta for the Claude Platform. The announcement describes a way for a long-running agent to send its conversation once and then transmit only the messages added since the previous response.
What the API keeps between turns
The first request starts a thread with the complete conversation. Later requests include new messages and the previous response ID. Anthropic says each thread is held encrypted in memory and expires five minutes after its last request. After expiry, the client resends the full conversation to start another thread.
This targets the time spent transferring a growing history of messages, tool results, images or documents. It does not mean the model stops reading earlier context. Claude still reads the full conversation on each turn, and token usage stays the same.
How to interpret the latency figure
With message threads enabled by default for most Claude Code users, Anthropic reports about a 35% reduction in median time to first token for requests carrying 500,000 to one million tokens. That is a result for a specified request-size range and latency measure, rather than a promise that every agent will finish its entire task 35% faster.
The company identifies active coding sessions, computer-use agents and document-heavy workflows as examples. Standard token prices and existing rate limits continue to apply.
What a client implementation needs
A useful integration should track the response chain, handle expiry and preserve a complete local conversation for restarting a thread. Measure time to first token separately from total task duration, especially when tools or generation dominate the elapsed time. The five-minute lifetime makes this most relevant to active sessions; it should not be treated as permanent conversation storage for jobs that resume much later.