Faster starts, unchanged context
Anthropic sees the feature helping coding and computer-use agents, along with workflows that process large numbers of documents or images. Those are settings where a conversation can grow long and each request may otherwise wait on the history before Claude starts responding.
The 35% figure is about time to first token, not total response time. The announcement does not say how the improvement was measured, or whether it holds across different conversation lengths and workloads.
The trade-off is still there
The encrypted history lasts only between requests and is removed after five minutes without activity. But storage is not the same as context reduction: Claude still reads the entire conversation, and token consumption does not change.
I think that makes Message Threads an infrastructure improvement, not a way to make long-running agent work cheaper. The unanswered question is how much of the wait comes from loading the history versus generating the response; without that breakdown, 35% is useful, but hard to translate into a workflow-level gain.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X