Anthropic Says Claude Now Writes 80% of Its Maker's Code
Anthropic's research arm just put a number on something everyone suspected but nobody had measured cleanly. As of May 2026, Claude authored more than 80% of the code merged into Anthropic's own production systems. Before the in-house coding agent rolled out in February 2025, that number was low single digits. Engineers there now ship roughly 8x as much code per quarter as they did across 2021 to 2025.
The report is about recursive self-improvement, the moment an AI system can fully and autonomously design and build its own successor. Anthropic is careful: we are not there yet, and it is not inevitable. But they argue it could arrive sooner than most institutions are ready for, and they back it with internal data instead of vibes.
The methodology is the interesting part. They pulled real Claude Code sessions from January to March 2026 where researchers worked with Claude on open-ended problems, like figuring out why a training run kept crashing. Then a separate Claude model judged whether the human or the AI had suggested the better next step. The AI was already winning a meaningful share of those calls. That is the part that matters. It is not about typing code faster, it is about who has the better idea about what to do next.
The honest read is this. When the company building the frontier model tells you its own development is being accelerated by its own model, and shows the receipts, the intelligence-explosion conversation stops being a thought experiment. Worth reading the full thing at https://www.anthropic.com/institute/recursive-self-improvement
← Back to all articles
The report is about recursive self-improvement, the moment an AI system can fully and autonomously design and build its own successor. Anthropic is careful: we are not there yet, and it is not inevitable. But they argue it could arrive sooner than most institutions are ready for, and they back it with internal data instead of vibes.
The methodology is the interesting part. They pulled real Claude Code sessions from January to March 2026 where researchers worked with Claude on open-ended problems, like figuring out why a training run kept crashing. Then a separate Claude model judged whether the human or the AI had suggested the better next step. The AI was already winning a meaningful share of those calls. That is the part that matters. It is not about typing code faster, it is about who has the better idea about what to do next.
The honest read is this. When the company building the frontier model tells you its own development is being accelerated by its own model, and shows the receipts, the intelligence-explosion conversation stops being a thought experiment. Worth reading the full thing at https://www.anthropic.com/institute/recursive-self-improvement
Comments