AnthropicAnthropic announced two changes intended to make Claude feel more responsive on October 16. Claude Code now gzips the conversation it sends on every turn, and cloud sessions start up as soon as the user begins typing.
Anthropic’s reported latency improvements. The first three rows show p90 wait excluding model time; the API row shows p50 time to first token including model time. Image: ClaudeDevs.
The changes address different parts of the wait before a response. Compression reduces the data sent with a conversation, while earlier session startup overlaps preparation with the time a user spends composing a request.
For coding agents, these delays can accumulate across repeated interactions even when the underlying model is unchanged. Improving transfer and startup behavior can therefore make a workflow feel faster without changing its reasoning capabilities.
The accompanying chart reports improvements from August to October 2026. At the 90th percentile, wait before the first token, excluding model time, fell from 4.07 to 1.33 seconds for local Claude Code sessions (3× faster), 16.99 to 6.00 seconds for cloud sessions (2.8×), and 3.36 to 2.17 seconds for Claude chat and Cowork (1.5×). The API figure uses a different measure: median time to first token, including model time, fell from 1.53 to 1.01 seconds (1.5×). These are Anthropic’s reported product-level improvements; the announcement does not isolate the contribution of gzip or earlier session startup.
Source: ClaudeDevs announcement.