Anthropic let an internal Claude model run a perf sprint on claude.ai: 3,000+ merged changes, 3.1x faster on average
A 23 September post on claude.dev describes an internal research model working from standing instructions in a Slack channel. It picked four user journeys from usage data, built deterministic benchmarks (Valgrind instruction counts, React commit counts, DOM mutations), and shipped PRs across 150+ concurrent threads with no customer-facing incidents. At P75, web fresh load dropped from 3,085ms to 550ms, desktop cold start from 6,310ms to 3,328ms, and sending a message in Cowork cloud from 928ms to 48ms. Techniques included a static HTML composer that accepts typing before React hydrates, V8 code-cache precompilation, and memoizing finished code blocks. The takeaway for builders is that deterministic benchmarks are what made autonomous optimization safe here.
↳ Follow the thread