Notes on working with AI
Message Threads: faster, not cheaper
Message Threads stop long agent conversations from re-uploading everything on each request. Anthropic reports faster first replies; the bill stays.
Every message re-sends the whole conversation
The Claude API remembers nothing between requests. To continue a conversation, your code sends the entire history again with every new message. In a long agent session at half a million tokens, that is hundreds of pages uploaded again just to add one sentence.
What changed
In early November 2026 Anthropic opened Message Threads in public beta. On the first request you send the full conversation as a thread of type create. From then on you send only the new messages, plus the id of Claude's previous response, as a thread of type continue. The server keeps the thread encrypted in memory for five minutes after your last request. After that it returns a 404, and you start again by sending the whole conversation, so you still keep your own copy.
The two numbers that matter
Anthropic reports a drop of about 35% in median time to first token on requests of 500k to 1M tokens, measured on Claude Code. That is Anthropic's figure, and I haven't seen an independent benchmark. The second number is that the bill doesn't move: Claude still reads the whole conversation, and you pay for those tokens. This saves waiting, not money.
Where it helps
A coding agent working across a large repository, where every step waits less. A support agent carrying a long case history. An analysis agent working through a set of contracts over one long session. Anywhere an agent loops many times over a big context. For short conversations you won't notice anything.
How to use it
The simplest route is the thread helper in the SDK, which adds the beta header (message-threads-2026-08-12) for you. By hand, add that header, send the first request as create, and send later ones as continue with the previous response id, without repeating the system prompt or the tools. Handle the 404 by resending. The feature isn't available for zero data retention or HIPAA configurations. Anthropic says Claude Code already uses it for most users, so there is nothing to switch on there.
If you only use the chat apps
The intuition carries over. Long chats are slower and use your limits faster. When a conversation gets long, ask Claude for a handoff summary and continue in a fresh chat.