Anthropic's team has identified signs of approaching recursive AI self-improvement: over 80% of code for current products is now generated by Claude, while engineer productivity has grown eightfold in a single quarter. Experts warn that if this trend continues, systems could autonomously design their successors before public institutions are ready for such challenges.

Metrics confirm the acceleration: in April, Claude completed 800 fixes that would have taken a human four years, and the success rate on open tasks reached 76%. The Mythos Preview model delivered a 52x speedup in AI training tasks, while in safety research, AI agents closed 97% of gaps compared to 23% by humans over a similar period.

Despite this progress, humans retain an advantage in strategic judgment, and human review may become the primary bottleneck in development speed. Anthropic acknowledges the potential need for a temporary pause in developing advanced systems to allow regulatory frameworks to adapt, but warns that without global coordination, unilateral slowdowns would merely advantage less cautious competitors.