How we made claude.ai 3x faster in two weeks / claude.dev Blog
The blog post How we made claude.ai 3x faster in two weeks
claude.ai web experience and Claude desktop app by an average of 3.1x.Key Highlights
- Four Core Journeys Targeted:
- Launching the app: Fresh load on web improved 5.6x (3,085 ms → 550 ms); desktop cold start improved 1.9x (6,310 ms → 3,328 ms).
- Starting a conversation: Web chat sped up 1.5x; desktop chat 2.1x; Claude Code desktop 2.4x.
- Loading a conversation: Web chat improved 2.4x; desktop chat 2.8x; Claude Cowork cloud 3.5x; Claude Code desktop 2.1x.
- Sending a message: Web chat sped up 3.1x; desktop chat 2.2x; Claude Cowork cloud 19x (928 ms → 48 ms); Claude Code desktop 4.8x.
- How They Did It:
- Autonomous AI Loop: Ran the sprint from a single Slack channel with an internal research model (Claude Tag) participating in every thread. Claude identified bottlenecks, ran benchmarks, submitted PRs, and monitored deploys.
- Deterministic Lab Benchmarks: Used Valgrind instruction counting (
node --predictable), V8 call counts, React commit metrics, style recalc counts, and layout shifts to measure changes without flakiness before deploying. - Key Technical Optimizations: Built a static HTML composer for instant typing during React init, precompiled V8 code caches, eliminated late sidebar layout shifts, optimized UTF-16 string handling for syntax highlighting, and optimized streaming renderers to hold 120 fps on 120 Hz displays.
- Guardrails & Safety: Merged over 3,000 changes using strict CI tests, automated reviews, short-lived feature flags (nearly 200 used), and human approvals without a single customer-facing incident or rollback.
