Anthropic's Message Threads Cut Time to First Token About 35% on Long Agent Runs

Anthropic’s message threads cut median time to first token by about 35 percent for requests of 500k to 1M tokens — a latency gain, not a cost saving. Anthropic announced the feature, now in public beta in the Messages API on the Claude Platform, so that agents stop resending their entire conversation on every request.

The mechanism is simple. The first request carries the full conversation and opens a thread through a new thread parameter. Later requests include only the new messages plus the ID of the previous response. Token usage does not change, because Claude still reads the whole conversation each turn, and the feature is billed at standard token rates. Threads are held in memory, encrypted, and never written to disk. A thread expires five minutes after its last request, after which the client resends the full conversation.

The problem it addresses is real for anyone shipping agents. A coding agent can make hundreds of requests in a single session, and once a conversation reaches hundreds of thousands of tokens, resending everything delays the start of Claude’s response. Anthropic says the feature is on by default for most Claude Code users, and names long-running coding agents, computer use agents that would otherwise re-upload screenshots, and document- or image-heavy workflows as the best fits. It is available on the native platform, Claude Platform on AWS, Claude in Microsoft Foundry and Google Cloud.

If you run long-lived agents, two practical checks follow. First, plan around the five-minute expiry: an agent that idles past it pays the full-resend delay once before reopening a thread. Second, measure time to first token on your own workload before and after. Anthropic’s 35 percent figure covers requests of 500k to 1M tokens, so shorter contexts should be expected to gain less.