Claude message threads: long agents stop re-uploading the whole conversation every turn
Every turn of a long agent run used to ship the entire conversation back to the API, screenshots and tool output included. Anthropic's message threads, in public beta since November 2, end that. The first request carries the full history. After that, the client sends only the new messages plus the previous response ID.
The effect shows up in latency, not the bill. Anthropic says median request size went from about 800 KB to 1.8 KB, and median time to first token dropped about 35% on requests between 500K and 1M tokens. It is already on by default for most Claude Code users. Token usage, pricing and rate limits do not change, because Claude still reads the full conversation every turn. This is a network and upload fix, not a cost cut.
Mechanics matter if you build on it. You pass a thread parameter with the beta header message-threads-2026-08-12. Threads live in memory, encrypted and never written to disk. A thread expires 5 minutes after its last request and is deleted within an hour of last use, after which the client has to resend everything. So any agent that pauses for a human approval or a slow tool call longer than five minutes still needs full-replay fallback logic. It works on the Claude API, Claude Platform on AWS, Microsoft Foundry and Google Cloud.
The pattern here is that agent workloads are now shaping API design. Chat-sized requests never needed this. Computer-use and coding agents dragging a million tokens of context around do, and the lab is redesigning the wire protocol around them.
Docs: https://claude.com/resources/articles/message-threads Release notes: https://platform.claude.com/docs/en/release-notes/overview
← Back to all articles
The effect shows up in latency, not the bill. Anthropic says median request size went from about 800 KB to 1.8 KB, and median time to first token dropped about 35% on requests between 500K and 1M tokens. It is already on by default for most Claude Code users. Token usage, pricing and rate limits do not change, because Claude still reads the full conversation every turn. This is a network and upload fix, not a cost cut.
Mechanics matter if you build on it. You pass a thread parameter with the beta header message-threads-2026-08-12. Threads live in memory, encrypted and never written to disk. A thread expires 5 minutes after its last request and is deleted within an hour of last use, after which the client has to resend everything. So any agent that pauses for a human approval or a slow tool call longer than five minutes still needs full-replay fallback logic. It works on the Claude API, Claude Platform on AWS, Microsoft Foundry and Google Cloud.
The pattern here is that agent workloads are now shaping API design. Chat-sized requests never needed this. Computer-use and coding agents dragging a million tokens of context around do, and the lab is redesigning the wire protocol around them.
Docs: https://claude.com/resources/articles/message-threads Release notes: https://platform.claude.com/docs/en/release-notes/overview
Comments