The Pulse
Anthropic Says Claude Helped Make Its Apps 3× Faster
Anthropic says a two-week sprint using Claude helped speed up claude.ai and the Claude desktop app, with measured gains across four common user journeys.
AI.info Team ·
Anthropic says Claude helped its engineers make claude.ai and the Claude desktop app about three times faster in a two-week sprint. The company published the account on September 23, describing a workflow in which Claude identified performance problems, built benchmarks, proposed code changes and monitored deployments while people set priorities and approved the work. Anthropic says the improvements covered four common user journeys that account for 95% of activity.
Three loading times that fell sharply
At the 75th percentile, Anthropic says the time from a fresh claude.ai load to a page where a user could type fell from 3.1 seconds to 0.55 seconds. Starting a new Claude Code session dropped from 0.8 seconds to 0.3 seconds, while loading a Claude Cowork cloud session fell from 2.6 seconds to 0.73 seconds. The figures describe selected journeys, not a claim that every operation in either app became three times faster.
Anthropic estimates that the changes save users tens of thousands of hours of waiting each day. The post does not provide independent measurements or a detailed breakdown of that estimate, so the overall speedup and time savings remain the company’s reported results. The team says it tracked thirteen measurements across web and desktop, with each starting at a user interaction and ending when the result appeared.
“[This model] is a numbers demon.”
— Shelley Vohr, an engineer in the channel
The team says it met twelve of the thirteen targets by day three. The initial sprint plan included about twenty projects, while additional opportunities emerged as Claude examined the measurements and suggested follow-up work.
Claude proposed fixes; people set the limits
Some changes targeted what users saw first. Anthropic placed a static version of the message composer in the page’s HTML so visitors could begin typing while React initialized. Other work kept the composer mounted during navigation, prefetched sessions when users hovered over them and cut sidebar re-renders by 90%.
The team also used Claude to investigate slower code paths. In one example, it profiled the routine that assembles a conversation’s message tree and a scanner for status lines in Claude Code output. Anthropic says instruction counts fell 48% and 31% on those paths, while their wall-clock times fell 78% and 44%, respectively.
Claude ran as an internal research model the company described as roughly comparable to Opus 5.5, using its Claude Tag product in beta. Engineers defined the goals, made tradeoffs and approved changes; Anthropic says more than 3,000 changes were merged without a customer-facing incident or rollback. The post describes the system as supervised work, not an autonomous performance overhaul.
Benchmarks became guardrails
Because milliseconds can fluctuate between runs, the team sought repeatable measurements that could run in the lab and catch regressions in continuous integration. It tested measures such as instruction counts, React commits, function calls, style recalculations and DOM mutations, discarding benchmarks that did not correspond to improvements in user-perceived latency.
When an optimization improved a benchmark, Anthropic says it added a threshold that future code changes could not exceed. The company also used feature flags and gradual rollouts: higher-risk changes went first to employees, then to 1% of users, before broader release. Those controls mattered when a layout shift appeared after the static composer shipped internally. The cause was Chrome resizing the page, not the composer handoff: Chrome’s managed-new-tab footer made the page shorter, then disappeared after the first paint, shifting the content as the page grew. Anthropic says the issue was latent and had not appeared before the page began painting so early; the team pinned the layout across the resize and added a test for the prerender flow.
The remaining work is beyond the headline number
Anthropic says the 95th percentile, other user journeys and very long conversations still need work. Its reported gains therefore mark progress on common measured tasks, rather than a finished performance target for every use of Claude’s apps.
The company’s account presents measurement as the engine of the sprint: once Claude had a benchmark to improve, it could test changes and repeat the cycle. Human reviewers still chose which problems mattered and whether a speed gain justified the complexity. Anthropic’s engineering post says the team plans to describe separate work that resulted in contributions to Electron, Chromium and Node.js.