In a two-week performance sprint, Anthropic engineers used an internal research model, Claude Tag (comparable to Opus 5.5), to accelerate the claude.ai web and desktop applications by approximately 3x. By integrating Claude into a dedicated Slack channel, the team established a continuous optimization loop where the AI identified bottlenecks via Datadog and other telemetry, built benchmarks, and proposed code improvements. The sprint targeted four user journeys comprising 95% of activity, resulting in significant latency reductions: 75th percentile time to a typeable page dropped from 3.1s to 0.55s, new Claude Code sessions improved from 0.8s to 0.3s, and Claude Cowork cloud sessions decreased from 2.6s to 0.73s. Technical optimizations included implementing a static HTML composer to enable typing during React initialization, precompiling V8 code caches for the desktop shell, and reducing sidebar re-renders by 90%. Claude also addressed low-level inefficiencies, such as reducing instruction counts in message tree assembly by 48% and optimizing syntax highlighting by converting UTF-16 strings to one-byte strings to avoid slower V8 regex paths. To ensure stability, the team utilized automated CI guardrails, feature flags, and incremental rollouts, merging over 3,000 changes without customer-facing incidents.