Eigen RadarAI
Analysis

Claude message threads let agents send only new messages after the first request

Anthropic has opened message threads in public beta for developers using Claude. An agent sends the whole conversation once, then appends new messages using the previous response’s identifier. The change reduces repeated uploads during long sessions. Claude still processes the full history, while prices and rate limits remain unchanged and temporary threads expire after inactivity.

Artificial Intelligence··Morning
Wide view of a server testing room, with memory modules exposed in a server on extended rails and network connections in the adjacent rack.

An agent sends the conversation once

Anthropic has made message threads available in public beta for the Claude application programming interface, which developers use to connect their software to its models. The feature changes how an AI agent sends its conversation history. Its first request contains the full conversation; subsequent requests can include only new messages and the identifier of Claude’s previous response. Earlier messages and tool results can remain attached to the thread instead of being uploaded again at each step.[1], [2]

Less uploading leaves full-context processing unchanged

A long coding session can involve hundreds of requests, with earlier messages and tool results growing to hundreds of thousands of tokens, the units used to process model input. Repeatedly transmitting that material adds to the wait before an answer begins. Anthropic says threads reduce that transmission delay. Claude still reads the complete conversation on every turn, so the change leaves token usage, standard prices and existing rate limits unchanged. Anthropic says it has enabled threads for most Claude Code users.[1]

The temporary thread expires after five minutes

Between requests, Anthropic holds each thread encrypted in memory without writing it to disk. The thread expires five minutes after its last request. A client then needs to send the full conversation again to start another thread, and developers must keep their own copy of the history. The beta does not qualify for zero-data-retention configurations. Anthropic describes the feature for active sessions, including coding, computer use and work with documents or images, where each new request builds on substantial earlier material.[1]

References

  1. News sourceThe Next WebClaude API message threads enter public beta↩1↩2↩3
  2. News sourceAnthropicAnthropic launches message threads in public beta for the Claude API↩