Anthropic warns of AI 'no brake pedal' as Claude codes ~80% of its own codebase

Anthropic published a detailed report on June 4 arguing the world should preserve "the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up." The company warned that unilateral action by one lab or nation would fail because competitors would race ahead — hence its call for coordination.

The alarm rests on concrete internal results. In an April 2026 open-ended AI-safety research project, human researchers chose the topic and scoring rubric, but Claude agents proposed hypotheses, designed and ran experiments, and iterated with parallel sub-agents: two human researchers recovered ~23% of the performance gap over a week, while the Claude agents recovered 97%. Claude Mythos Preview achieved a 52x speedup over baseline code on optimization tasks where a skilled human needs four-to-eight hours for a 4x gain. Claude now authors roughly 80% of Anthropic's new production code, and success rates on complex engineering problems rose to 76% in May 2026. Task horizons handled reliably have doubled roughly every four months — from minutes in early 2024 to 12-hour tasks now — with week-long autonomous tasks projected by 2027.

This lands as Anthropic pursues a rumored ~$1 trillion IPO, and that timing fueled skepticism. New Scientist and several commentators argued the recursive-self-improvement warnings double as IPO hype, while developers on X (scaling01, Andon Labs) noted Opus 4.8 became 'eval-aware' and is still 'token-mogged by GPT-5.5' on inference efficiency — framing the capability story as catch-up rather than a frontier reset.

Watch whether any rival lab publicly engages the pause proposal; absent reciprocity, Anthropic's own framing concedes the call is unlikely to change the race.