Claude: Messages API Can Now Retain Conversations Server-Side Between Turns
On November 2, Anthropic launched a public beta of “message threads” for Claude's Messages API: after an initial full payload, an agent only transmits new messages and the previous response ID. Pricing and token consumption remain unchanged, but time to first token drops on long sessions.
Anthropic launched on November 2, 2026, a quiet yet significant public beta feature for agent developers: message threads for Claude’s Messages API. Until now, this API was strictly stateless: on every turn, the client had to resend the entire conversation, including every tool result, file, and screenshot. With threads, the conversation remains on the server for a short time, and the client only sends what is new.
How It Works
The mechanism relies on a single parameter, thread, added to standard requests.
- Create. The initial request is sent just like today, with the entire conversation, system prompt, and tools, plus
"thread": {"type": "create"}. The API stores the exchange under the response ID. - Continue. Subsequent turns send only the new messages and
previous_message_id, the ID of the last response. The system prompt and tools are no longer resent; certain parameters (model, thinking, output format, beta headers) must remain identical, or the API will return an error. - Replay. The service operates on a “best-effort” basis: a thread expires 5 minutes after its last request and is deleted at most 1 hour after its last use. If the API cannot resume it, it returns a 404 error, and the client must resend the complete conversation. The official SDKs offer a helper that handles this retry automatically.
The feature requires the message-threads-2026-08-12 beta header. It is available on the Claude API, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude on Google Cloud (managed by Anthropic).
What Changes, and What Stays the Same
The improvement is in latency, not cost. Anthropic already uses it in Claude Code, where threads are enabled by default for most users.
| Metric published by Anthropic (Claude Code) | Value |
|---|---|
| Median time to first token, requests between 500,000 and 1 million tokens | approx. −35% |
| Median request payload size in long conversations | from approx. 800 KB to 1.8 KB |
| Pricing and rate limits | unchanged |
Figures provided by Anthropic, not independently verified.
Claude still re-reads the entire conversation on every turn: token consumption remains identical, and prompt caching is not automatic—it must still be explicitly requested. What threads eliminate is resending, validating, and tokenizing the entire history with every request. Anthropic highlights long-running coding agents, computer-use agents (where screenshots stay in the thread), and heavy document- or image-processing workflows.
Key Limitations
The documentation is transparent about the trade-offs. The client remains the source of truth: it must retain its own copy of every message, as responses do not confirm whether a turn was successfully stored. A thread only supports append operations; modifying a message, summarizing history, or swapping tools requires starting a new thread. Threads do not work with Message Batches, token counting, or on-demand compaction, and no API exists to list, read, or delete a thread.
Regarding data handling, content is held in memory, encrypted, without writing to durable storage. However, the feature is not eligible for zero data retention (ZDR) and remains unavailable to HIPAA-enabled organizations. Anthropic has not shared a date for general availability.
Our take
This marks the partial end of a long-standing model API dogma: server-side statelessness. OpenAI had already taken this step with its Responses API; Anthropic is proceeding cautiously, keeping retention windows very short and leaving the client in charge of history. For developers building agents that accumulate hundreds of thousands of tokens, the benefit is concrete and free: shorter wait times before the first response and drastically smaller network payloads. Yet this is not a cost-saving measure, since tokens are billed identically, and error handling becomes more nuanced (expiry, replays, locked parameters). Teams bound by strict zero-retention requirements will have to sit this out for now.
Sources
- Claude Docs — Message threads Primary source
- Claude — Make long-running agents faster with message threads Primary source
- ClaudeDevs on X — public beta announcement (November 2, 2026) Primary source
- ClaudeDevs on X — benchmark metrics on Claude Code (November 2, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
