How we made claude.ai 3x faster in two weeks

A two-week engineering sprint made the core claude.ai web and desktop experiences roughly three times faster by focusing on four high-impact journeys (app launch, starting a conversation, loading a conversation, and sending a message) that account for 95% of user activity. At the 75th percentile, a fresh web load went from 3.1s to 0.55s, starting a Claude Code session fell from 0.8s to 0.3s, and a Claude Cowork cloud session dropped from 2.6s to 0.73s; thirteen measurements across platforms averaged a 3.1x speedup (geometric mean). Engineers merged over 3,000 changes with no customer-facing incidents by running work in a single Slack channel where an internal Claude Tag (beta) research model guided prioritization, benchmarking, and monitoring.

The team converted vague latency problems into deterministic metrics and CI guardrails - using Valgrind instruction counts for Node hot paths and browser measures like V8 call counts, React commit counts, layout/style recalcs, and DOM mutations. Claude found and fixed hot paths (e.g., resolving message IDs once instead of three times) and proposed targeted changes: a static composer to allow typing during React init, precompiled V8 cache, keeping the composer mounted, prefetching sessions on hover, and reducing sidebar re-renders by 90%. Lab wins tracked to real user time (example: a 48% instruction reduction produced a 78% wall-clock speedup), and the iterative loop - trace, benchmark, PRs behind flags, monitor, ratchet - became the operating model for further gains.