Anthropic proposes metrics for the pace of AI development inside frontier labs

Anthropic proposed a set of public measurements for how fast AI is being developed inside frontier labs, and published its own figures against them. The company framed the metrics as a way to make the abstract call to pace the frontier — which Dario Amodei had set out days earlier — into something an outside observer could actually track, rather than a claim labs simply ask to be trusted on.

Three measurements were offered. The first indexes how much of a lab’s own AI research is led by AI: on a scale from AL0 to AL5, Anthropic reported that Claude “leads” 26% of its AI research and development work as of August 2026, up from under 1% in February, with no tasks yet running at full autonomy. The second covers oversight of the roughly 30,000 research agents Anthropic said it runs at once — online monitors that block dangerous actions in real time, at a reported blocking rate of about 0.002%, backed by offline review after the fact. The third tracks how compute is split: during the measured week, about 6% of the compute going to AI research went to safety work, rising to about 12% of the compute spent specifically on AI-driven research.

Anthropic acknowledged the numbers are its own and not independently verified, and presented the framework as a starting point it hoped other labs and outside groups would adopt or contest. The proposal extends the transparency theme running through the month’s pacing debate: where Amodei’s essay argued for slack between capability and control, this attaches candidate yardsticks to it.