Anthropic Calls for Pause on Frontier AI Development

Artificial intelligence stalwart Anthropic called for a global pause in frontier model development by invoking fears that the technology has outpaced society’s ability to keep up.

See Also: How AI Increases the Risk of Enterprise Data Exposure

A pause “would be good for the world,” said co-founder Jack Clark and Anthropic Institute head Marina Favaro – so long as it doesn’t let “the least cautious actors catch up technologically.”

Sparking the company’s appeal to humanity is concern that the time when AI can create its own successor is coming soon. As of May 2026, more than 80% of the code written and accepted into Anthropic’s software was authored by Claude, up from the low single digits before Claude Code launched in early 2025. The typical engineer is producing eight times as much code per day as in 2024, not because they are working harder, but because Claude is doing most of the writing while engineers direct and review.

The authors said that Claude-written code was worse than human-written code as of late 2025, roughly at parity today and expect it to be better within the year. A similar pattern is emerging for research, they said.

The authors described a test in which Claude was asked to suggest the next step in a real research investigation. Anthropic’s best model beat the human’s choice 51% of the time in tests conducted last fall. By April, that had risen to 64%.

The term the authors use for where this leads is recursive self-improvement. Anthropic said AI has not reached that point yet and that it may not be inevitable. But, “it could come sooner than most institutions are prepared for,” Clark and Favaro said.

Hence the plea for a coordinated global pause, requiring all leading AI labs in multiple countries to stop frontier AI development under the same conditions, with the means to verify that others had done so too. No such mechanism exists and building one would be harder than control regimes for other technologies, which took decades to coalesce – and still aren’t universally followed across the globe.

Not everyone reads the data as pointing toward recursive self-improvement. Cognitive scientist Gary Marcus said in a blog post describing impressive coding tools but not much more. “All they have really shown is just faster coding, entirely under human control,” Marcus wrote. “A faster coding tool will probably not end the world.”

Critics have said that Anthropic is calling for the brakes in a race in which it is a front runner, having filed for an initial public offering and with a valuation of $965 billion.