Anthropic Says Claude Now Writes Over 80% of Its Merged Code — but Insists It's Not Recursive Self-Improvement Yet

In a new essay titled "When AI builds itself," Anthropic disclosed that more than 80% of the code merged into its production codebase as of May 2026 was authored by Claude — up from the low single digits before Claude Code launched in research preview in February 2025. "Claude is now writing Claude," as chief product officer Mike Krieger put it, capturing how quickly the company's own software has become a product of its own model.

The productivity numbers are striking, if imperfect. In the second quarter of 2026, Anthropic says the typical engineer merged roughly eight times as much code per day as in 2024 — though the company is unusually candid that this figure "is almost certainly an overstatement of the true productivity gain," since lines of code measure volume, not value. A cleaner signal: Claude's success rate on the most open-ended engineering problems hit 76% in May 2026, a 50-point jump in just six months.

Leadership has been pushing the story further. Speaking at the Cisco AI Summit, Krieger said that "for most products at Anthropic it's effectively 100% just Claude writing," with the model also acting as a demanding code reviewer that flags security holes and suggests refactors. Boris Cherny, the creator of Claude Code, reportedly writes 100% of his own code through Claude and has not hand-edited a line since November 2025. The trajectory validates a prediction CEO Dario Amodei made in March 2025 — that AI would write 90% of code within three to six months, and "essentially all" of it within a year — that many dismissed as hype at the time.

Yet Anthropic is deliberately careful about what this does and does not mean. "We are not there yet, and recursive self-improvement is not inevitable," the essay states, identifying judgment as the missing ingredient: large gaps persist when it comes to Claude choosing which goals to pursue, rather than executing ones humans set. Writing code is not the same as deciding what to build, and that gap is what still separates a fast assistant from a system that improves itself. Still, the company warns the moment "could come sooner than most institutions are prepared for," and calls for verification infrastructure and the ability to pause frontier development — adding bluntly that "we don't have that long" to build it.

The disclosure also lands against a backdrop of mounting concern about AI-written code in the wider industry, where adoption is far lower — Microsoft has cited around 30% and Google over 25% AI-generated code. Security researchers warn of an "illusion of correctness," and one survey found that roughly one in five CISOs had experienced a breach tied to AI-generated code. Anthropic's pitch is that the same model that writes the code can also review it more rigorously than a rushed human would — a claim that, like the recursive-self-improvement question itself, will be settled by results rather than slogans.