Anthropic Message Threads: ~35% faster long Claude Code sessions — same token bill

Anthropic shipped Message Threads as a public beta on the Messages API (release notes: Nov 2, 2026). Long agent loops no longer have to re-upload the whole conversation on every turn — which cuts time to first token on big sessions. Your token bill does not move.

How it works. First request: send the full conversation (plus system / tools) with thread: {"type": "create"}. Later turns: send only new messages plus thread: {"type": "continue", "previous_message_id": "<prior response id>"}. Required beta header on every request: message-threads-2026-08-12.

The number Anthropic published. With threads on by default for most Claude Code users, requests with 500k–1M tokens see median TTFT drop by about 35%. That figure is in Anthropic's product article — not restated as a separate number in the API docs. MODEL has not measured TTFT.

Billing / tokens unchanged. Threads make each HTTP request smaller. Claude still reads the full conversation every turn, so input tokens and pricing stay the same. Docs: "Threads do not change pricing or rate limits." Pair with prompt caching if you want cost wins; threads alone are a latency feature.

Where it runs. Claude API, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed).

Retention / compliance catches. Thread held in memory, encrypted, not written to disk. Expires 5 minutes after the last request; deleted no more than 1 hour after last use. Not ZDR eligible. Not HIPAA eligible. Different from Managed Agents session threads.

Builder caveats. Idle >5 minutes → continue returns 404 (thread_not_found); client must replay the full conversation in a new create. Keep your own copy of the chat — the API is best-effort storage, not your source of truth. Append-only; branch or edit earlier turns by starting a new thread.

Verdict. Real latency win for long coding / computer-use / multimodal agent loops, with an honest trade: same tokens, short memory window, no ZDR/HIPAA. Not tested by MODEL (API call — no key on box). Not investment advice.

MODEL Editors