Anthropic launches Messages API message threads beta for faster long agents
On 2 Nov 2026 Anthropic put message threads into public beta on the Claude Platform Messages API: after one full upload, later turns send only new messages to cut time-to-first-token. Anthropic says it is on by default for most Claude Code users, with about a 35% median TTFT drop on 500k–1M token requests, and unchanged pricing and rate limits.
Verified on Anthropic’s article Make long-running agents faster with message threads and the same-day Claude Platform release notes (November 2, 2026): Anthropic launched message threads in public beta for the Messages API.
Anthropic states that long agent loops often resend the full conversation and tool results every turn, which adds delay before Claude starts responding once context is large. With message threads, an app sends the full conversation once (including system and tools), then on later turns sends only new messages plus the previous response id; the API rebuilds the conversation before calling the model. Requests use the message-threads-2026-08-12 beta header and a thread parameter.
Anthropic says message threads is on by default for most Claude Code users, and that requests with about 500k to 1M tokens see median time to first token (TTFT) fall by about 35%. Between turns, threads are held in memory, encrypted, and not written to disk; a thread expires about 5 minutes after its last request, after which the client resends the full conversation to start a new one. Token usage is unchanged because Claude still reads the full conversation each turn; standard token rates and existing rate limits apply.
Anthropic highlights long-running coding agents, computer-use agents (screenshots can stay with the thread), and document/image-heavy workflows. Availability covers the native Claude Platform plus Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed). Docs also note limits such as no Message Batches support, no on-demand compaction with threads, and that threads are not ZDR-eligible.
Figures and capability claims are Anthropic’s; they are not independent IdeaDawn retests.