Anthropic has published a set of proposed measurement tools aimed at illuminating the pace of AI development inside frontier labs. The company argues that as AI systems grow more powerful and begin to automate more of the process of building themselves, the public needs more information about how models are produced.
The post organizes measurement around three areas: the extent to which AI is building the next version of itself rather than being built by humans; the ability to oversee and intervene in actions AI agents take on Anthropic's systems; and the resources that power the development of more capable models. Anthropic says it is also sharing a snapshot of these metrics from inside the company, while noting that the numbers would likely shift if there were coordination on pacing the frontier, as called for by CEO Dario Amodei.
On AI-led R&D, Anthropic describes a prototype Anthropic R&D Automation Index, built by cataloguing every kind of AI R&D task at the company and rating how automated each currently is using Epoch AI's Automation Level scale, which runs from AL0 (no AI involvement) to AL5 (full autonomy). According to the findings, Claude is not operating fully autonomously for any measured subset of AI R&D work, but it "leads" 26% of that work, and the share of work at or above "AI collaborates" is above 90%.
For agent oversight, Anthropic reports that as of August 2026 there were approximately 30,000 agents doing research and engineering work at any one time on its most-used internal platform. It tracks three metrics: coverage, review latency, and escalation rate. The company says individual agents rarely misbehave in its monitoring data to date, but warns that when millions or billions of agents operate in the economy, even rare events can happen regularly.
On compute allocation, Anthropic examined a snapshot of how it used all of its compute from July 13 to July 20, sorting workloads into categories and asking how much of the compute going to AI R&D was safety work. It cautions that safety research tends to use less compute than frontier training runs, making compute an imperfect proxy for safety focus.
Anthropic says it plans to embed independent third-party evaluators from multiple organizations, giving them access to internal processes, systems, and data comparable to what internal risk assessment teams have. These third parties would verify safety practices, report incidents, and monitor key metrics.