Anthropic puts numbers on how much of AI development is now done by AI

Marina Favaro and Jack Clark, writing for the Anthropic Institute, argue that recursive self-improvement is not a future event but a process already underway, and try to put numbers on it from inside Anthropic. The doubling rate of task complexity Claude can handle has shortened from about seven months to about four. A year ago the company's models could carry out four-minute coding tasks; one year later they were running twelve-hour ones. More than 80 percent of code merged into Anthropic's codebase in May 2026 was authored by Claude, and engineers shipped roughly eight times as much code per quarter as in 2024. The authors are careful that lines of code is a soft metric, but they pair it with concrete incidents: Claude debugged a crash affecting tens of thousands of training jobs in about two hours, work that normally takes two to three days.

The research-side numbers are the more interesting ones. On a code-optimization benchmark Claude went from a 3x speedup a year ago to about 52x in April 2026. On a 129-case test where Anthropic researchers had to make a non-obvious call between research directions, the model beat the human choice 64 percent of the time, up from 51 percent six months earlier. In one open project, agent runs recovered 97 percent of the gap toward solving an AI safety problem the team had been working on. The piece ends by sketching three scenarios: the trend stalls and current capabilities just diffuse, compounding efficiency continues with humans setting direction, or the loop closes and humans move to oversight. Favaro and Clark put the middle case as the most likely and call for research now on verifiable mechanisms to pause AI development if a country or company decides it needs to.

Why it matters

If the doubling rate is really four months, the people deciding what oversight looks like have less calendar time than they probably assume, and Anthropic publishing internal numbers is a useful pressure to make rivals share theirs. For engineers, the more grounded read is the 80 percent code-authoring figure: that is what a working agent loop looks like in practice, and it sets a bar for what to expect at your own company within a year.