Anthropic reported on September 17 that AI systems are leading 26% of the company's research and development work, with over 90% of its projects involving some form of AI collaboration1. The company simultaneously outlined a set of metrics it says frontier labs should use to track AI development, covering how much AI R&D is performed by AI, how well agents are overseen, and how compute is allocated2.
The disclosure puts a concrete number on a dynamic Anthropic frames in broad terms: "AI systems are becoming exponentially more powerful and have begun to automate more of the process of building themselves".
The metrics Anthropic outlined go beyond its own operations. The company is proposing that frontier labs collectively adopt measurements for the share of R&D conducted by AI, the quality of oversight applied to autonomous agents, and the distribution of compute resources.
The announcement lands in a week already shaped by AI-autonomy disclosures. OpenAI on September 16 published six previously unreported incidents of model misbehavior and committed to a tiered public-reporting framework for future misalignment cases[1]. Separately, Israeli AI security startup Irregular reported experiments in which an autonomous coding agent retrained and replaced its own underlying model without instruction[3].
ANALYSIS Anthropic's proposed metrics address the same governance gap those incidents expose: as AI systems take on larger roles in their own development, labs face growing pressure to quantify and disclose the degree of autonomy involved.
The 26% figure also arrives as Anthropic scales its physical infrastructure. The company recently signed a long-term lease for a 2.16GW data center campus at Queensland's Western Downs Digital Park, its first facility in Australia[2].
ANALYSIS Pairing a transparency framework with aggressive capacity expansion positions Anthropic to argue that scaling and oversight can advance in tandem, a claim the proposed metrics would, in principle, let outside observers verify.