Anthropic: AI Now Writes 80% of Our Code, Pushing Toward Self-Improving Systems

Anthropic is publishing internal data showing how much of its own AI development is now being done by AI. As of May 2026, Claude authored more than 80% of code merged into Anthropic’s codebase, up from low single digits before Claude Code launched in February 2025. Engineers are merging roughly 8x as much code per day as they did in 2024, with research staff self-reporting around 4x productivity gains under Mythos Preview. Anthropic concedes lines-of-code is a noisy metric and likely overstates true uplift, but argues the inflection points are unmistakable.

The piece frames this as steady progress along a curve that ends in recursive self-improvement, where AI systems autonomously design and train their successors. External benchmarks back the trajectory: task-length capability is doubling roughly every four months, SWE-bench has saturated, and CORE-Bench (replicating published research) went from 20% success to saturation in fifteen months. Anthropic says Claude can already execute well-specified engineering and research tasks at or above human level, but still lags on choosing what to work on, which is the gap separating today’s models from a system that could build its own replacement.

The authors do not claim recursive self-improvement is inevitable or imminent, but warn it could arrive faster than institutions expect. If models capable of building their successors emerge, the difficulty of monitoring, securing, and steering them rises sharply, which is the policy point the piece is really pushing.