How Anthropic Made Claude.ai 3x Faster in 14 Days

If you use Claude on the web or through its desktop apps, you might have noticed everything feels a lot snappier lately. Anthropic has now detailed how a team of engineers spent two weeks overhauling the core user experience across claude.ai, the desktop app, and Claude Code.

On average, the platform is now roughly three times faster across key daily interactions, saving users thousands of hours of collective wait time every single day.

What makes this speed sprint unusual is how the engineers pulled it off. Instead of manually auditing thousands of lines of code, the team assigned Claude itself to manage performance bottlenecks, write benchmarks, draft pull requests, and monitor production deployments.

Anthropic started by looking at real user monitoring data to identify where people spent the most time waiting. They narrowed the focus to four primary user journeys that make up roughly 95% of all activity on the platform:

1. Launching the app: Fresh loads on the web and cold starts on desktop.
2. Starting a conversation: Creating new chat sessions across web and terminal tools.
3. Loading a conversation: Opening existing threads and cloud workspaces.
4. Sending a message: The client-side work required to process user input.

Before the sprint, launching a fresh session on claude.ai took over three seconds to reach a typeable state at the 75th percentile. Following the optimization push, that same page load takes just 0.55 seconds, an 82% speed improvement.

Similarly, loading a conversation on the desktop app dropped from 1.35 seconds to under half a second, while sending a message on client-side cloud sessions sped up by a staggering 95%, dropping from 928 milliseconds down to a nearly instantaneous 48 milliseconds.

During the two-week push, the team merged more than 3,000 code changes. Despite the rapid pace and massive volume of updates, Anthropic reports that the sprint was completed without a single customer-facing incident or emergency rollback.

For users, the main takeaway is a noticeably faster web and desktop experience. For the engineering team, it proved that when an AI model can measure a problem, it can systematically fix it.