Anthropic Engineering Team Uses Claude Model to Speed Up Web and Desktop Apps Threefold

Anthropic engineers increased the performance of the claude.ai web interface and the Claude desktop application by approximately three times during a two-week sprint conducted in August, according to a technical overview published on the company's developer blog (https://claude.dev/blog/how-we-made-claude-ai-faster/) and shared on Hacker News. The team executed the latency-reduction project primarily within a dedicated Slack environment, relying on an internal research model comparable to Claude Opus 5.5 to identify application bottlenecks, write benchmarks, generate software patches, and track continuous deployments.

The engineering effort targeted four primary user workflows that account for 95% of customer activity across 13 distinct performance metrics on web and desktop platforms. At the 75th percentile, initial load time to a typeable page on claude.ai dropped from 3.1 seconds to 0.55 seconds. Session creation within Claude Code fell from 0.8 seconds to 0.3 seconds, while initialization of cloud sessions in Claude Cowork decreased from 2.6 seconds to 0.73 seconds. Anthropic estimated that the combined latency improvements save users tens of thousands of wait hours on a daily basis.

The optimization sprint was structured around an automated Slack interface operating alongside human developers. Using a Datadog Model Context Protocol (MCP) server, the AI agent—designated as Claude Tag (beta)—analyzed operational data and executed tasks across more than 150 active Slack threads. Over the course of the sprint, the team merged more than 3,000 changes, reaching peak volume on days when over 200 changes landed in the codebase. Anthropic reported zero customer-facing outages or code rollbacks throughout the initiative.

Early structural changes targeted immediate startup and navigation delays. Engineers embedded a static input composer into the application's base HTML so users could begin typing while React initialized, alongside precompiling a V8 code cache for the desktop app's main process. Additional updates maintained composer mounting across conversation transitions, prefetched session data when users hovered over navigation items, and reduced sidebar re-rendering frequency by 90%.

To validate code changes rapidly without relying on noisy wall-clock measurements, the team established deterministic benchmarks within continuous integration pipelines. Using Valgrind with predictable execution flags in Node.js, the AI system tracked instruction counts, React commit counts, V8 function execution, and DOM mutations. On critical execution paths—such as the routine assembling conversation message trees—eliminating redundant dictionary lookups reduced instruction counts by up to 48%, driving down real-world latency by 78%.

Expanding the automated profiling revealed several non-obvious technical bottlenecks. The system discovered a single CSS selector adding 24 milliseconds to every DOM change, 6,900 React hooks and 900 store subscriptions firing on individual keystrokes, and a legacy location reload call causing 500,000 hidden daily reloads. In another instance, non-Latin-1 characters like em dashes forced V8 to store strings in UTF-16 format, triggering a one-second page freeze during syntax highlighting until the team introduced a one-byte string conversion step.

The sprint also addressed visual layout shifts by tracking data through the browser Layout Instability API, correcting rendering sequences that previously caused post-load visual movements on 31% of web page loads. To mitigate risk across the accelerated development schedule, engineers deployed nearly 200 short-lived feature flags to control rollout cadences, clearing out more than half of the temporary flags before the conclusion of the two-week project.