Claude now writes most of the code at Anthropic

Hook

The company building the frontier model is now mostly running on the frontier model. The engineers set the goal; Claude writes the code.

The news

On June 4 Anthropic published "When AI builds itself" and put a number on it: more than 80% of the code merged into Anthropic's own codebase in May 2026 was authored by Claude, up from low single digits before Claude Code launched in research preview in February 2025. The typical engineer now merges 8x as much code per day as in 2024, with Claude writing it while the human directs and reviews. Anthropic says human code review has become the new bottleneck, and adds that it would be good for the world to have the option to slow or pause frontier development if rival labs verifiably did the same. Read it here.

On your desk

Strip the recursive-self-improvement headlines and the operator lesson is plain: the people who build Claude don't write Claude line by line anymore. They aim it. They review it. The work moved from typing to directing. That is the same posture you take as a CEO with a coding agent and a harness, and Anthropic just confirmed it scales to the hardest software on earth. If the engineers closest to the model have handed it the keyboard, the question isn't whether you should. It's how much of your own work still runs through your hands out of habit.

What you should do next

Pick one task you still do by hand, hand it to Claude this week, and if you've never built with it, start with your first software tool.