According to CoinMeta, Anthropic is in public beta within its messages API launch message threads (message threads), with the aim of reducing the overhead associated with the repeated uploading of conversation history by AI agent. In the past, coding agent required the re-sending of previous messages and tool execution results with each call to Claude; however, now only the complete conversation needs to be uploaded once, with subsequent messages and replies to the previous round being sent individually. The server temporarily saves the historical records, eliminating the need for repeated transmissions, verifications, and tokenization. According to Anthropic, among 500,000 to 1 million token requests, the median waiting time for the first token has been reduced by about 35%, and the median size of a single request for long conversations has decreased from approximately 800kb to 1.8kb, representing a reduction of about 99.8%. Nevertheless, the model still needs to process the complete context, so the token usage and costs remain unchanged. The message threads saved by the server will expire after 5 minutes of inactivity, at which point the historical records need to be re-uploaded.