Anthropic: Claude "leads" 26% of its AI R&D

Anthropic says AI systems are becoming exponentially more powerful and have begun to automate more of the process of building themselves. As the world considers slowing the pace of frontier AI development, the company argues the public needs more information. That framing sets up the piece not as a capability boast but as a transparency exercise: it offers measurement tools for how much AI is building its next version, how well Anthropic can oversee and intervene in agent actions, and where compute goes.

The three measurements connect to Anthropic's broader policy proposals. Its Responsible Scaling Policy risk reports already publish capability evaluations, while the Advanced AI Framework proposes transparency obligations governments could require. The measurements here focus on the production process, so they can be correlated with those capability evaluations. The body also notes Anthropic plans to embed independent third-party evaluators from multiple organizations with access comparable to internal risk teams, and says external red-teaming has happened before, with METR independently testing the offline monitoring platform.

The stakes hinge on verifiability and comparability. Anthropic acknowledges obstacles to cross-lab comparison, including the lack of a common methodology and its use of its own models as judges. The compute snapshot also covers one week, which the body says shows the measurement can be made but not a meaningful trend. How much outside parties can trust these numbers, and whether other developers adopt similar reporting, is likely to determine their influence.