Anthropic message threads cut time to first token by 35% in long chats

Anthropic message threads are in public beta: the Messages API feature lets a long conversation skip the full upload on every turn, The Decoder reported on 3 November. Anthropic says it cuts the median time to first token by about 35% for requests of 500,000 to 1 million tokens.

Until now, an agent had to resend the whole conversation with every request. The API then validated and tokenized all of it again before Claude could start answering. In a coding session that grows with each tool result and file edit, that overhead grows too, and you wait longer for the first word as the context window fills.

How a thread works

• Create: you send the first request as usual, with the full conversation, system prompt and tools, and add a thread field set to create.
• Continue: on later turns you send only the new messages and the ID of the previous response, and leave out the system prompt and tools.
• Replay: a thread expires 5 minutes after its last request. A continue request then returns a 404 error, and you send the full conversation again to start a new thread.

Anthropic holds the thread in memory, encrypted under a server-managed key, and never writes it to durable storage. It deletes the thread no more than an hour after its last use. For that reason the feature is not eligible for zero data retention. Your application stays the source of truth, and the documentation tells you to keep your own copy of every message and build the replay path from the start.

What you pay

Billing stays the same. Claude still reads the full conversation on every turn, so token usage, standard rates and rate limits do not change. Only the request payload shrinks. You can still turn on prompt caching for each turn.

Who gets it

The 35% figure comes from Claude Code, where Anthropic turned threads on by default for most users. Anthropic expects the biggest gains for a long-running AI agent that writes code, for computer-use agents that would otherwise re-upload screenshots, and for workflows heavy on documents and images.

The beta works on the Claude Platform, Claude Platform on AWS, Microsoft Foundry and Anthropic-managed Claude models on Google Cloud. The documentation lists support across current models, from Claude Fable 5.1 and Opus 5.5 back to Sonnet 4.5. Developers enable it with the message-threads-2026-08-12 beta header. The release comes weeks after Anthropic's IPO filing on Nasdaq.

The 35% figure is Anthropic's own median for Claude Code traffic; your latency gain depends on how long your conversations get.