On Thursday, Anthropic released an essay titled “When AI builds itself,” publicly calling on frontier AI labs worldwide to slow down or pause model development in a coordinated, verifiable manner. The essay disclosed internal company data for the first time: as of May this year, over 80% of the code merged into Anthropic’s codebase was generated by Claude, and per-engineer quarterly code output had increased roughly 8 times compared to the 2021–2024 baseline. The essay, co-authored by Marina Favaro, head of the internal research institute, and Jack Clark, head of policy, issued a core warning: AI is accelerating toward the tipping point of “Recursive Self-Improvement” — when systems can autonomously iterate on next-generation models without human intervention — which “may arrive before most institutions are ready.”
The essay cited multiple external benchmarks and internal experiments to support the trend of acceleration: the time limit for AI to independently complete tasks roughly doubles every four months; Claude Opus 4.6 can currently handle tasks requiring 12 hours of sustained work stably, and is expected to cover complex work spanning weeks by around 2027; Claude Mythos Preview achieved roughly a 52x speedup in internal code optimization tests, compared to only about 3x a year ago (roughly 4x for a skilled engineer); in the most complex open-ended tasks, Claude’s success rate reached 76% in May this year, an improvement of roughly 50 percentage points in six months. Clark estimated that the full tipping point for recursive self-improvement could arrive within two years.
The essay likened the proposed global slowdown mechanism to nuclear weapons treaties, but frankly acknowledged that the difficulty of verification far surpasses previous efforts: “It took decades to build the infrastructure and trust for a verification regime, and we don’t have that much time”; a unilateral pause would only change the frontrunner rather than establishing the broad deliberative mechanism needed. Anthropic plans to convene dedicated discussions with policymakers, researchers, and other AI companies in the coming months. The essay’s release timing is sensitive: Anthropic just completed a $65 billion financing round at a valuation of nearly $965 billion, and this week filed a confidential IPO document. External skeptics question its motives: venture capitalist and Trump advisor David Sacks accused it of pursuing a “regulatory capture agenda”; others view it as a marketing tactic to promote its flagship model Mythos; Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, commented that while the essay contains elements of both self-reflection and marketing, Anthropic’s central assessment of AI’s trajectory is sincere and worth paying attention to.