The Anthropic Institute has put numbers on something it has hinted at for months: more than 80% of the code merged into its production codebase in May 2026 was written by Claude, up from low single digits before Claude Code launched in February 2025. The typical engineer now merges 8x as much code per quarter as in 2024.
The sharper data point is on research, not engineering. In one internal study, Claude-powered agents tackled an open AI-safety problem and recovered 97% of the available performance gap over 800 cumulative hours and about US$18,000 in compute; two human researchers managed 23% in roughly a week. Anthropic is careful to say this is not recursive self-improvement, the point at which a system autonomously designs its successor, and that it is not inevitable. Humans still set the direction.
What makes the piece notable is the ask. Co-authored by policy lead Jack Clark, it argues the world should build the option to slow or pause frontier development, including verification systems so labs can confirm rivals have actually stopped. That is a frontier lab publishing its own acceleration metrics and then arguing for a brake. Whether anyone can build a credible one in time is the open question.