📅 November 3, 2026 • claude.com
What happened
Anthropic has released Message Threads in public beta on the Messages API, changing how agents talk to Claude. Instead of resending the entire conversation on every request, an agent sends the full history once to start a thread, and every following request carries only the new messages plus the ID of the previous response. The feature is already enabled by default for most Claude Code users. Per Anthropic, median time to first token dropped about 35% on requests carrying 500K-1M tokens, and the median request size in long conversations fell from roughly 800 KB to 1.8 KB — about 99% less data on the wire. Both figures come from Anthropic's own Claude Code rollout, not from independent tests.
How it works
Between turns the thread is held in memory on Anthropic's servers, encrypted and never written to disk, and it expires five minutes after the last request — the client then resends the full conversation to start a new one. Crucially, this is a latency feature, not a cost feature: Claude still reads the complete conversation on every turn, so token usage, billing and rate limits stay unchanged. Message threads work across all Claude models in beta on the native Claude Platform, AWS, Microsoft Foundry and Google Cloud. Developers calling the API directly pass a thread parameter plus a beta header, while the official SDKs include a helper that manages the thread and replays the conversation automatically. There are beta-era limits worth knowing: per Anthropic, threads are not eligible for zero data retention, are off-limits for HIPAA-enabled organizations, and there is no API to list, read or delete a thread — so your own copy of the conversation remains the store of record.
IMPACT
For teams running long agent sessions — coding agents, computer-use workflows, document-heavy pipelines — this attacks exactly the bottleneck that grows with context size: the wait before the model starts answering. The savings scale with context, so the bigger your agent's working memory, the bigger the win. Thinner requests also cut upload time on constrained networks — relevant for agents shuttling screenshots and documents back and forth. The five-minute expiry makes threads a fit for active sessions rather than background jobs, and the missing ZDR support rules them out for regulated data. The practical pattern: adopt threads for interactive agent loops, keep a full-history replay path in your own code as the fallback, and treat the request-size reduction as an infrastructure win that leaves both your token bill and your architecture intact.
Sources: claude.com, thenewstack.io, the-decoder.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)