After telling the world, AI will kill you; now Anthropic wants companies to follow these three ‘metrics’ to know what AI models can do |
Anthropic CEO Dario Amodei called for the pace of development of AI models to slow down last week. In a latest, the company has now laid out measurement tools for companies to highlight three critical aspects of AI development: “The extent to which AI is building the next version of itself, as opposed to being built by humans; Our ability to oversee and intervene in actions that AI agents take on Anthropic’s systems; The resources that power the development of more capable models”.“As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows,” Anthropic said in the blog post adding “This means better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information.”
Anthropic wants to measure how much AI is building AI
The first metric tracks AI-led AI research and development. Anthropic said it has created an AI R&D Automation Index to measure how much of its AI research and development work is performed by Claude. The company uses a scale from AL0 to AL5. At AL3, AI “collaborates” with humans. At AL4, AI “leads” the work and can complete most of a task from a high-level instruction. AL5 means the AI works fully autonomously without a human in the loop.As of August 2026, Anthropic said Claude was not fully autonomous for any measured AI R&D work. However, Claude was at the AL4 “leads” level for 26% of Anthropic’s AI R&D work, while more than 90% of the work was at AL3 or above.Anthropic said the share of work at or above the “AI collaborates” level had risen from below 1% in February 2026 to the current level for the measured work.
Anthropic tracks how AI agents are monitored
The second metric looks at oversight of AI agents. In the blog post, Anthropic said about 30,000 agents were carrying out research and engineering work at any one time on its most-used internal platform in August.The company said 100% of these agents’ actions pass through an online monitor before they are executed. Their actions are also ingested by an offline monitoring system after they happen.Anthropic analysed more than 1 billion decisions made by its research and engineering agents during August. About 0.002%, or roughly one in 47,000 decisions, were blocked by the online monitor.The offline system flags about 100,000 transcripts a week for further review. Around 50 of the highest-priority cases each week are escalated to human reviewers.
Anthropic measures how much computing power goes to AI safety
The third metric tracks compute allocation, or how Anthropic uses its computing resources. The company examined its compute use over one week and divided workloads into categories, including AI research and safety work.Anthropic said about 6% of compute used for AI R&D during the period was allocated to safety work. For compute used specifically for AI-driven AI R&D, the figure was about 12%.The company said these are conservative estimates and that compute share is not a perfect measure of how much work is being done on safety.Anthropic wants other AI labs to publish similar numbersAnthropic said other frontier AI developers could publish the same types of measurements.“As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows,” the company said.It added: “This means better measuring the development of AI, reporting on it publicly, and giving society an opportunity to decide how to use this information.”
