Anthropic has called for the ability to slow or temporarily pause frontier AI development, citing the rapid acceleration of AI systems designing newer, more capable successors with minimal human involvement. The company highlighted that over 80% of the code merged into its production codebase is now written by its AI model Claude, a significant increase from low single digits before the launch of Claude Code in 2025, according to medianama.com.
In a recent paper, Anthropic explained that AI is increasingly accelerating its own development internally. The typical engineer at Anthropic merged eight times as much code per day in the second quarter of 2026 compared to 2024. The company noted that human engineers are shifting roles from performing technical tasks to reviewing AI outputs and deciding which problems to address, as the actual execution of tasks now requires minimal human time.
Anthropic's paper also detailed the growing capabilities of its Claude models, which have advanced from handling software tasks taking minutes in 2024 to those lasting several hours and more recently up to 12 hours. This progress underscores the potential for recursive self-improvement, where AI systems autonomously design and develop their successors, raising concerns about managing the pace of frontier AI development.
The company’s call for a mechanism to slow or pause AI development aims to provide a safeguard as AI systems become more autonomous. Anthropic’s internal benchmarks and public data highlight the increasing complexity and duration of AI-driven tasks, emphasizing the need for oversight in this rapidly evolving field.