Anthropic releases Message Threads public beta on Claude platform, cutting time to first token about 35% for very long conversations

Anthropic has released Message Threads as a public beta on the Claude platform, and also through AWS, Microsoft Foundry and Google Cloud. Instead of resending the whole conversation with every API request, only the changes since the last message are sent. The thread is held encrypted in memory between requests and expires after five minutes of inactivity. Anthropic says time to first token falls by about 35% for requests of 500,000 to 1 million tokens. Token usage does not change, because Claude still reads the full history.

This mainly helps developers running very long sessions, such as coding agents, computer-use agents and tools that work through many documents or images, where each turn currently resends a huge payload. Faster first responses could make those tools feel more responsive. It does not cut cost or add capability, so it is not a reason to change vendors or budgets. The five-minute expiry means workloads with gaps between turns may not benefit, and teams will need to handle expired threads in their code.

The 35% figure is Anthropic's own claim and applies only to the extreme 500,000 to 1 million token range, so it should not be assumed for typical conversations. The headline also says nothing about total response time or cost, and no independent benchmark has been published. The coverage is a secondary report with no direct links to the documentation in the extracted evidence.

1. 01Independent latency measurements from developers across different context lengths (under 100,000 tokens and over 500,000), and whether the roughly 35% holds outside Anthropic's own test.
2. 02Whether the feature leaves beta, and any changes to the five-minute expiry, pricing or limits, plus how consistently it behaves on AWS, Microsoft Foundry and Google Cloud.
3. 03Whether rivals such as OpenAI or Google announce similar incremental-request or session-state features, which would show this is becoming a standard expectation.