How We Made claude.ai 3x Faster in Two Weeks
claude.dev Blog
Claude.ai got about 3x faster in a two-week sprint. Anthropic says Claude helped find the bottlenecks, and they shipped 3,000+ changes without a customer incident.
Based on reporting by claude.dev Blog — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
In August, Anthropic spent two weeks doing one thing: making claude.ai and the Claude desktop app feel much less sluggish. The company says the core experience ended up about three times faster, after users complained it was slow and, as the team put it, they were right.
The work ran out of a single Slack channel, with Claude in every thread. The team started by focusing on four journeys that made up 95% of user activity, then built benchmarks for thirteen measurements across web and desktop. At the 75th percentile, a fresh load of claude.ai to a typeable page fell from 3.1 seconds to 0.55. Starting a Claude Code session dropped from 0.8 seconds to 0.3, and loading a Claude Cowork cloud session from 2.6 seconds to 0.73.
Anthropic also used Claude Tag, an internal research model it says is roughly comparable to Opus 5.5, to hunt down bottlenecks, build benchmarks, ship fixes, and watch deploys. The team says Claude estimated the impact of projects in milliseconds, then helped prove those estimates in the lab before anything went live. That mattered because the sprint moved fast: twelve of thirteen initial targets were hit by day three, and the original project list kept expanding as Claude found more work.
A lot of the wins were very specific. The team baked a static composer into HTML so users could type while React initialized. It precompiled a V8 code cache for the desktop shell. It kept the composer mounted between conversations, prefetched sessions on hover, and cut sidebar re-renders by 90%. Elsewhere, Claude found a message-tree routine and a Claude Code status-line scanner whose instruction counts could be cut, then turned those into CI ratchets after wall-clock time moved with them.
The scale is what stands out. More than 3,000 changes merged in two weeks, with no customer-facing incident or rollback. And the team didn’t just ship faster code; it built a loop where every measurement became a target, every target became a benchmark, and every benchmark became a guardrail.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI people should actually pay attention to: not flashy demos, but a model grinding through dull engineering work until the numbers move. The hype crowd wants sentient co-pilots; the real win is a ruthless little performance goblin that can count, test, and keep score.
Read more about this at: claude.dev Blog