How a two-week sprint made Claude.ai three times faster
Anthropic's web team spent two weeks in late August tearing into Claude.ai's speed, and the writeup by Raymond Wang, Sam Attard, and Issac G. is a clear account of how they did it. The headline results are large: a fresh Claude.ai load at the 75th percentile fell from 3.1 to 0.55 seconds, roughly 5.6x, the desktop cold start nearly halved, and Claude Code session start dropped from 0.84 to 0.35 seconds. Across four core user journeys, which cover about 95 percent of activity, they claim a 3x improvement and tens of thousands of user-hours saved every day.
The method is the interesting part, and it is summed up in one line: once Claude can measure something, it can make it faster. Rather than chase noisy wall-clock timing, the team built deterministic benchmarks using instruction counts, V8 call counts, React commits, and DOM mutations as targets the model could hill-climb against, then checked those lab wins against real field data. In one case, cutting CPU instructions by 48 percent tracked a 78 percent wall-clock speedup. The work ran as a loop across more than 150 Slack threads: identify a slow interaction, have Claude build a benchmark and trace it, ship a small change behind a feature flag, verify after deploy, then lock the win in with a ratcheting benchmark so it could not regress.
Guardrails kept three thousand merged changes from causing incidents: unit tests, automated review with human approval, staged rollouts, and layout-shift and keystroke tests to catch visual or data-loss regressions. One fix was almost comic, the post notes: em dashes in code blocks triggered UTF-16 handling that slowed syntax highlighting by a full second, fixed in twenty lines.
Why it matters
If you own a web app's performance, the transferable idea is the ratcheting deterministic benchmark: pick a proxy metric a model or CI can optimize without flaky timing, then prevent regressions automatically. It also shows agent-driven optimization works at scale when the measurement and rollout safety are solid, which is the hard part to copy, not the model itself.