Anthropic Report Reveals Accelerated AI Self-Improvement, Calls for Global Development Pause

Anthropic’s latest internal report highlights a significant acceleration in AI development, largely driven by its own frontier models like Claude. The company reveals that its engineers are now shipping eight times more code per quarter compared to 2021-2025 levels, with over 80% of merged code in its codebase being Claude-authored as of May. This exponential growth is attributed to models autonomously running code and working over long time horizons, leading to substantial productivity multipliers across engineering and research teams. Claude has demonstrated superhuman capabilities in tasks such as code optimization, achieving a 52x speedup compared to a 4x gain by skilled human researchers. Furthermore, Claude-powered agents have successfully conducted open-ended research, recovering 97% of a research gap over 800 cumulative hours, and current models are now outperforming human judgment in steering research sessions 64% of the time. The report suggests that the human role in AI development is narrowing, shifting primarily to direction-setting and judgment as AI assumes more of the execution.

These advancements bring Anthropic to seriously consider the implications of “recursive self-improvement” – an AI takeoff where systems improve themselves beyond human comprehension. While acknowledging the potential for enormous good, Anthropic also expresses significant concerns about the risks of losing control over increasingly capable AI systems. In a notable stance, the company advocates for the option to temporarily pause or slow frontier AI development, akin to a nuclear peace treaty, to allow societal structures and alignment research to catch up. However, they emphasize that such a pause would only be effective if implemented globally and verifiably by all major frontier labs, to prevent less cautious actors from gaining a competitive advantage. Anthropic plans to organize discussions with policymakers, researchers, and other AI companies to explore coordination mechanisms and address the profound questions raised by the rapid pace of AI progress.