Anthropic adds server-side message threads to cut agent latency

Summary

• Anthropic launched message threads in beta on its Messages API on November 2, 2026.
• The API keeps conversations server-side, so follow-up requests carry only new messages instead of full history.
• Anthropic says threads lower time-to-first-token in long agent loops without changing pricing or rate limits.

The latest

Long conversations and agent loops on Anthropic's Messages API can now send smaller requests, after the company added message threads in beta on November 2, 2026 under the message-threads-2026-08-12 header. According to Claude platform release notes, the API retains the conversation server-side and each subsequent request carries only the new messages. Anthropic says the feature does not change pricing or rate limits.

Details

• How it works: According to the documentation, a thread is opened by sending the full conversation with system, tools and a thread field of type create. Every later turn sends only new messages with thread type continue and previous_message_id, dropping system and tools from the request body.
• The stated purpose: Anthropic says the goal is faster and more consistent time-to-first-token in long agent loops. Without a thread, each turn re-uploads and re-tokenizes the entire conversation, a cost that grows with every additional round of an agent's execution.
• Availability: Threads are offered on the Claude API, the Claude platform on AWS, Claude models on Google Cloud under Anthropic management, and Claude in Microsoft Foundry. The beta applies across all models, with documentation examples built on claude-opus-5-5.
• The field rules: System and tools are creation-only; sending either in a continuation, even as null, returns a 400 error. Model, tool_choice, thinking, mcp_servers excluding the auth token, context_management, output_config, container skills, top-level cache_control and the anthropic-beta set must match the creation request or the API returns 400 thread_fingerprint_mismatch.
• What stays flexible: Per-turn fields that may vary include max_tokens, stop_sequences, stream, metadata and service_tier. The API stores neither MCP server URLs nor their tokens nor the container identifier, so those are resent each time. Changing system or tools requires a new thread.
• Lifetime and failure: A thread expires 5 minutes after the last request, with each stored response resetting the count, and is deleted within at most an hour of last use. Threads are best-effort: a failed continuation returns 404 with error_code thread_not_found, obliging the application to resend the full conversation.
• Caching limits: There is no automatic caching. A top-level cache_control in each request caches the last cacheable block, with a ceiling of 4 cache_control points per request including the top-level one, and system and tools points counted again on each continuation. Exceeding the ceiling returns 400 thread_unsupported_request.
• Other constraints: Threads are append-only and each response identifier can be continued once, so editing or branching requires a new thread. Streaming delivers the identifier in message_start but continuation is permitted only after message_stop, and no thread continues after a refusal stop_reason.
• SDK helper: Anthropic says client.beta.messages.thread adds the header, manages creation and continuation, strips system and tools on continuation, retries once on thread_not_found, thread_already_continued or thread_continue_in_flight, and opens a new thread after a refusal. It covers Python, TypeScript, C#, Go, Java, PHP and Ruby.
• Not supported: Message Batches, token counting, which returns 400, compaction and on-demand compaction, context editing other than clear_thinking_20251015 with keep all, HIPAA-readiness and inference-hooks organizations. Threads are not eligible for ZDR or HIPAA, are held encrypted in memory under a server-managed key, and never written to durable storage.

Background

Claude platform release notes list adjacent changes: Anthropic disabled fast mode for Claude Opus 4.8 on October 30, 2026, with removal set for November 30 and users directed to Claude Opus 5.5; launched a Usage and Cost Analytics API in beta on October 29; and raised the maximum Messages API request size from 32 to 128 megabytes on October 28.

Between the lines

The design reads as latency engineering rather than state management. Anthropic offers no interface to list, read or delete threads, caps their life at 5 minutes of inactivity, and labels them best-effort, requiring applications to keep their own copy of every message. Compaction is also unsupported, with a summary-based new thread offered as the workaround. The server holds the transcript only long enough to skip re-tokenizing it.

What's next

The beta header message-threads-2026-08-12 marks the feature as provisional, so watch for general availability, support for token counting and compaction, and whether the fast mode removal lands on November 30, 2026.

Source: Anthropic — Claude platform release notes and documentation