Anthropic puts message threads in public beta so Claude agents send only new messages each turn
The Messages API now holds a conversation in memory for five minutes between requests; Anthropic says median time to first token falls about 35% on 500k–1M token Claude Code requests, while token usage and prices stay the same.
Why it matters
Coding and computer-use agents can make hundreds of requests in a session, and until now each one re-sent every earlier message, tool result and screenshot. Message threads keep that history on Anthropic's servers, encrypted in memory and not written to disk, so later requests carry only what is new. Because Claude still reads the full conversation every turn, the change cuts waiting time rather than cost. Watch how developers handle the five-minute expiry and when the feature leaves beta.
Key facts
- Public beta on the Claude Platform from November 2, 2026
- Median time to first token down about 35% on 500k–1M token requests
- On by default for most Claude Code users
- Thread expires 5 minutes after its last request
- Token usage, rates and rate limits unchanged
- Also on Claude Platform on AWS, Microsoft Foundry and Google Cloud