Anthropic says a 2‑week sprint made claude.ai 3.1 times faster
Engineers credit a model working in a single Slack channel, and put a fresh web load at 550 milliseconds against 3,085 before the sprint.
Anthropic said on Wednesday that a two-week sprint made claude.ai and its desktop app 3.1 times faster, a geometric mean across 13 measurements at the 75th percentile. According to the post, the company compared real user monitoring on 13 August with 27 August, across four journeys it said make up 95% of activity.
A fresh load of claude.ai fell from 3,085 milliseconds to 550, the post said. A desktop cold start went from 6,310 milliseconds to 3,328, loading a conversation on the desktop app from 1,353 to 488, and sending a message in Claude Cowork from 928 milliseconds to 48.
The three authors, Raymond Wang, Sam Attard and Issac G., said the sprint ran in one Slack channel with Claude in every thread. An engineer would open a thread on a slow stretch, Claude would trace it and build a benchmark, then return with pull requests sized for review, anything user-visible behind a flag.
One example: message-tree assembly was resolving the same message ID three times. Cutting that removed 48% of the CPU instructions on the path and 78% of the wall-clock time, the post said, and the team then made any pull request that raised the count fail its tests.
The numbers are Anthropic's own and nobody outside the company has checked them. The post said the work targeted the 75th percentile, and that the 95th percentile, other journeys and the longest conversations may still have room to improve. Anthropic estimates the gains could save tens of thousands of user-hours of waiting a day.