Which figure stands today
Anthropic wrote that as of August 2026, on Epoch AI's automation rungs, Claude 'led' 26 percent of its AI R&D work and 'collaborated' on more than 90 percent. No measured slice was fully autonomous. The same text says about 30,000 agents were doing research and engineering on its most-used internal platform at any one time.[1]
The compute slice is narrower: in the week of 13-20 July, about 6 percent of compute that went to AI R&D was classed as safety work, and about 12 percent of compute that went to AI-driven AI R&D. Work that advanced safety and capability equally was excluded. The company sorted one week of jobs into its own bins. The text does not give a grid megawatt.[1]
Monitoring is not delivered capacity
An online monitor screens 100 percent of those agents' actions before they run; of more than a billion August decisions, about 1 in 47,000 were blocked. Humans are said to review blocked actions within one week. The figure is an internal gate count. A connected transformer and an energized rack are absent from this text.[1]
Embedding independent third-party evaluators with access comparable to internal risk teams is described as a plan. The snapshot that stands today is the agent count, monitor coverage and compute share.[1]