Anthropic significantly speeds up Claude

Anthropic has significantly sped up Claude.

In Claude Code, the wait time before the first token is generated has dropped by 3x. In chat, Cowork and the API, it is 1.5x faster.

First, Claude Code now sends the model the dialogue history as a compressed archive, which travels over the network faster.

Second, previously the cloud environment started only after a message was sent. Now it begins warming up as soon as you start typing, so it is already ready by the time the prompt is sent, eliminating extra startup time.

The new Claude Code version also adds idle compaction: if a conversation of 200k+ tokens has been idle for about 54 minutes, it is automatically compressed.

This matters because the prompt cache lasts for one hour. Compaction is performed shortly before the cache resets so the next message does not recompute the entire long context at full price (changelog).