Anthropic discusses the rise of self-improving AI

Anthropic has published a report examining recursive selfimprovement, the idea that future AI systems could help build and improve newer versions of themselves with less human involvement. The company says that process is not fully here yet, but it believes AI development is already speeding up in meaningful ways. The report points to internal coding data showing that Claude wrote most merged code at Anthropic in May, while engineers were producing far more code per day in 2026 than they were in 2024. Anthropic argues that these trends suggest AI is increasingly contributing to the systems that will shape later models. The company also said it is considering how the industry should respond if AI progress reaches a point where coordinated limits or pauses become necessary. OpenAI has raised similar concerns in recent policy work, and both companies say the risks around selfimproving systems and related safety issues deserve more discussion among labs and policymakers.