When AI Builds Itself: Anthropic's Progress Toward Recursive Self-Improvement

Anthropic has published an assessment, produced by the Anthropic Institute, of its progress toward "recursive self-improvement" — the point at which an AI system can autonomously design and develop its own successor. The authors say the industry is not there yet and that the outcome is not inevitable, but argue it could arrive sooner than most institutions are prepared for.

Evidence from inside the company

The most concrete figure is internal: as of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude. Before Claude Code launched in a research preview in February 2025, that share was in the low single digits. In the second quarter of 2026 the typical engineer merged about eight times as much code per day as in 2024, with humans directing and reviewing rather than typing. Anthropic cautions that lines of code is an imperfect measure of real productivity gains.

In April 2026, Claude shipped more than 800 fixes that reduced a class of API errors by a factor of a thousand — work the overseeing engineer estimated would have taken a human four years. An automated Claude reviewer now reads changes before they merge; a retrospective analysis found it would have caught roughly a third of the bugs behind past claude.ai incidents before production.

Evidence from public benchmarks

The article also cites METR's finding that the length of tasks AI systems can reliably complete on their own has been doubling roughly every four months, up from an earlier trend of seven. In March 2024, Claude Opus 3 handled software tasks that take a human about four minutes; a year later Claude Sonnet 3.7 managed tasks of about an hour and a half; a year after that, Claude Opus 4.6 handled 12-hour tasks.

SWE-bench, which gives a model a real open-source codebase and a real bug report, went from low single-digit scores to saturation in two years. Anthropic's own recurring test — asking Claude to make a small training script faster while passing the same correctness checks — moved from an average ~3x speedup in May 2025 to ~52x by April 2026.

The gap that remains

The piece is explicit about what has not been automated. Claude can be handed an underspecified engineering problem and work out how to solve it, and it can match or outperform skilled humans at executing a well-specified experiment. Large gaps persist, however, in choosing goals — deciding which problems are worth working on. That judgement separates today's systems from one that could design its own successor.

The authors sketch three futures: the trend stalls while today's capabilities diffuse widely; labs keep seeing compounding efficiency gains with humans setting direction, which the evidence is said to point toward; or AI systems become capable of full recursive self-improvement. They add that a credible slowdown would need verification that other labs had genuinely stopped, and say the Anthropic Institute will help build the systems such a pause would require.