Claude threads shrink long requests from 800 KB to 1.8 KB

Anthropic on November 2 opened message threads in public beta on the Messages API. In Claude Code, which has them on by default for most users, median time to first token on 500k-1M token requests fell about 35%. Long-running agents gain speed and pay the same bill, because Claude still reads the full conversation every turn and token usage stays the same. Zero-data-retention organizations can’t use threads at all.

The first request carries the full conversation and the thread parameter. After that, each request sends only the new messages and the previous response’s ID, Anthropic’s post says. The docs name the beta header message-threads-2026-08-12 and say system and tools stay out of continue requests. Each thread lives in server memory, encrypted and never written to disk. It expires 5 minutes after its last request, a continue then returns a 404 thread_not_found, and the client resends everything. The beta runs on the Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud (Anthropic managed), at standard token rates and existing rate limits.

ClaudeDevs on X puts the median long-conversation request at 1.8 KB, down from about 800 KB. Neither figure says whether the bytes are raw or compressed, and Claude Code has gzipped large request bodies since 2.1.300 on October 14. So the drop is a request size, not a measured network saving. Both headline figures come from Anthropic’s own client.

A thread doesn’t cache anything for you. Prompt caching still has to be requested, with at most 4 cache_control breakpoints per request, and the top-level field counts as one of them. Threads don’t work with Message Batches, token counting, compaction, or context editing other than clear_thinking_20251015 with keep: "all". They’re also unavailable to organizations that have HIPAA readiness or inference hooks turned on.