Only the difference sent from the second request on
Anthropic on Nov. 2 (Nov. 3 Japan time) added "message threads" in public beta to its developer-facing Claude API. Because the API holds the conversation on the server side between exchanges, developers need only send the new message and the ID of the previous response instead of resending the entire conversation each time. That shortens the time until Claude begins answering, the company said.
The first request sends the entire conversation, the system instructions and tool definitions, and specifies "create." From the second request on, developers specify "continue," attach the ID of the previous response and send only the new message. The system instructions and tool definitions from the first request carry over.
Request size down from 800KB to 1.8KB
According to Anthropic, in long Claude Code tasks the median size of each request fell from about 800KB to 1.8KB. In long conversations of 500,000 to 1 million tokens, the median time until the response begins fell by about 35%. In short conversations of a few thousand tokens, there is little difference, the company said.
Pricing and usage limits are unchanged. Because Claude rereads the entire conversation each time, input token pricing is also the same. What decreases is the amount of data transmitted, not the amount read. To lower costs, developers still need to specify prompt caching themselves, as before.
Deleted after 5 minutes; keep history locally too
Stored conversations expire 5 minutes after the last request and are deleted within an hour at the latest. When a conversation expires, it must be rebuilt by resending the whole conversation, so apps also need to keep the history. Conversations are encrypted and held in memory and are not written to disk, the company said.
Messages once sent cannot be edited, and the model, reasoning settings and other options cannot be changed midway. The feature cannot be used with the Message Batches bulk processing, under zero data retention (ZDR) agreements or under HIPAA terms for healthcare. Besides Anthropic's direct API, it is available on Claude Platform on AWS, the Anthropic-operated endpoint on Google Cloud and Microsoft's Foundry; Amazon Bedrock is not listed.
A run of API revisions
In October, Anthropic released the first new version of the Claude API in three years (related article) and raised the limit on data sent per request to 128MB (related article). It continues to revise the API for agents that handle long tasks.
