Anthropic engineers say claude.ai got 3x faster in two weeks by baking the composer into HTML, caching V8 code and cutting sidebar re-renders 90 percent

Anthropic engineers Raymond Wang, Sam Attard and Issac G. published a post on claude.dev on 23 September 2026 saying claude.ai and its desktop and cloud siblings got roughly three times faster across the last two weeks. The number is a geometric mean of 13 measurements taken across four user journeys, and the post lists both the numbers and the specific changes that produced each one.

The five journeys and the before-after numbers

The post reports a claude.ai web fresh load moving from 3,085 milliseconds to 550, a desktop app cold start from 6,310 to 3,328, a Claude Code desktop launch from 837 to 347, a web chat message send from 180 to 59, and a Claude Cowork cloud load from 2,566 to 728. On the fresh load that is a 5.6x improvement; on the send it is 3.1x; on the geometric mean across all 13 tracked measurements it is 3x.

What they actually changed

The post names the mechanism behind each win. On the web fresh load, a static version of the message composer is now baked directly into the HTML, so a user can start typing before React finishes initialising. On the desktop shell, the main process now warms a V8 code cache during install, so start-up does not repay the compile cost every time. On repeat sends, the composer is kept mounted between conversations, so React does not rebuild it. Session data prefetches on hover, and the sidebar's re-render count is down 90 percent from removing work that was firing on every keystroke. The post reports a 48 percent cut in CPU instruction count in one hot path after the team found and removed a duplicate message-id lookup.

Two lower-level changes are worth naming because they are the kind of thing a review does not usually catch. The team fixed cases where V8's hidden-class shape had gone megamorphic on dictionary lookups the code ran often, and cases where routine string handling was doing extra work because a common character encoded as a UTF-16 surrogate pair rather than a single code unit. Between those two, the post reports main-thread code blocking on a common send path down from 1.0 second to 0.35.

How they measured it, which is the part most teams skip

The post is unusually specific about the measurement stack. Deterministic benchmarks were built with Valgrind under the --predictable flag, so a change either lowered the CPU instruction count or it did not; the team also tracked React commit counts, style recalculations and DOM mutation counts as first-class metrics rather than wall-clock latency only. Real user monitoring came from Datadog, and every change deployed behind a feature flag with automated review and a human approval gate. A 120 Hz frame-budget monitor watches the reply-streaming loop, and a Layout Instability API integration flags a janky sidebar that would otherwise be reported as "feels slow" and never chased.

For a team already sitting on a large React application with the same shape as claude.ai, the useful move this week is to copy the measurement discipline, ahead of any one fix. An instruction count under a deterministic harness gives a change either a clean pass or a clean fail, which is the guardrail that lets a two-week push actually publish a 3x number without a rollback in the fourth week.

Source

Raymond Wang, Sam Attard and Issac G., claude.dev, How we made claude.ai 3x faster in two weeks, 23 September 2026.