Anthropic let an internal Claude model rewrite claude.ai's performance

Anthropic ran an internal Claude research model against its own product for weeks, and the model made claude.ai faster. Working from standing instructions in a Slack channel, it picked four user journeys straight from usage data, built its own deterministic benchmarks, and shipped more than 3,000 merged changes across 150+ concurrent threads. No customer-facing incidents, per Anthropic's writeup.

The numbers are the kind you'd expect from a team, not a script. At the 75th percentile, web fresh load dropped from 3,085ms to 550ms. Desktop cold start went from 6,310ms to 3,328ms. Sending a message in Cowork's cloud environment fell from 928ms to 48ms. Average speedup across the four journeys: 3.1x.

What actually moved those numbers is specific and boring in the best way. A static HTML composer accepts keystrokes before React finishes hydrating. V8 code-cache precompilation skips re-parsing JavaScript on repeat loads. Finished code blocks get memoized instead of re-rendered. None of that is novel engineering. What's new is who found it and verified it.

The benchmarks are the actual story here. Valgrind instruction counts, React commit counts, DOM mutation totals, these are objective, replayable numbers, not "the app feels snappier." That's what let an autonomous model touch production code across 150 threads without a human reviewing every diff for regressions. Take away the deterministic harness and the same setup is just a model committing code nobody can verify against a real baseline.

Anthropic doesn't say how much human review happened before merge, or whether this approach generalizes past performance work into correctness-sensitive changes. Those are different guarantees, and the post doesn't claim to have tested them.