Anthropic · 2026-09-23 · notable
How Claude made claude.ai 3x faster — 3,000 changes in two weeks
Anthropic engineers used Claude in a Slack channel to make claude.ai and the desktop app 3.1x faster across 13 measurements. More than 3,000 changes shipped in two weeks with no customer-facing incident or rollback.
Anthropic let Claude hunt performance bugs in its own apps for two weeks, and the speed gains were large and measured.
What is it?
Anthropic's claude.dev engineering blog describes an August 2026 sprint that made claude.ai and the Claude desktop app 3.1x faster on average (geometric mean of 13 measurements). The work was done by Claude Tag (beta), running an internal research model roughly comparable to Opus 5.5, with engineers in the same Slack channel.
How does it work?
The key step was measurement. Claude built deterministic lab benchmarks — CPU instruction counts under Valgrind, React commit counts, V8 call counts, layout-shift tracking — so each fix could be checked quickly. With 150+ threads running at once, changes shipped behind about 200 short-lived feature flags and rolled out from employees to 1% to all users. Examples include a static composer you can type in before React loads, and moving code blocks to single-byte strings before regex highlighting.
Why does it matter?
Fresh claude.ai loads went from 3,085 ms to 550 ms at p75, and sending a message in Claude Cowork went from 928 ms to 48 ms. The post also finds bugs normal metrics missed: 31% of page loads moved content after they became usable, although the page scored well under the layout-shift threshold. The authors' lesson: 'With Claude, measuring something makes it tractable.'
Who is it for?
frontend and performance engineers