Long-running agents stop resending full history each turn: median time to first token falls about 35% on 500k–1M-token requests
Synopsis
Anthropic has opened message threads in public beta on the Claude Platform, letting agents send only what is new since the last turn instead of resending the whole conversation history. Previously a coding agent could make hundreds of requests in one session, each carrying every earlier message and tool result, and once a conversation reached hundreds of thousands of tokens that resend added time before Claude began responding. With message threads on by default for most Claude Code users, requests of 500k to 1M tokens see median time to first token drop by about 35%, while Claude still reads the full conversation each turn, so token usage stays the same.
Interpretation
Long-running agents no longer need to resend the entire conversation history every turn; sending only what is new since the last turn continues the same conversation. Previously a coding agent might make hundreds of requests in one session, each resending every earlier message and tool result, and once a conversation reached hundreds of thousands of tokens that resend added time before Claude started responding. With message threads on by default for most Claude Code users, requests with 500k to 1M tokens see median time to first token drop by about 35%.
Use means adding the thread parameter to Messages API requests: the first request carries the full conversation and starts a thread, and each later request carries only the new messages and the ID of the previous response. Between turns each thread is held in memory, encrypted and not written to disk, and expires 5 minutes after its last request, after which the client resends the full conversation to start a new one. The feature is in public beta on the native Claude Platform and available on Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed), at standard Claude token rates and under existing rate limits.
Perspective
The demonstrated scope is message threads in public beta on the Claude Platform, covering the native platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed), at standard Claude token rates and under existing rate limits. It targets latency-sensitive active sessions, including long-running coding agents, computer use agents, and document- and image-heavy workflows.
The official description states that Claude still reads the full conversation on every turn, so token usage stays the same and the saving is waiting time from resending, not billed tokens. A thread expires 5 minutes after its last request, and the client then resends the full conversation to start a new one. The latency benefit grows with the context an agent carries, and the reported roughly 35% drop corresponds to the 500k to 1M-token request range.