Anthropic wants you to know Claude leads a quarter of its research, but "lead" doesn't mean what you think
decoder ·
Anthropic has published metrics on how much of its AI research is handled by models, how its agents are monitored, and how much compute is devoted to safety work. As of August 2026, Claude was classified at Epoch AI's AL4 level for 26 percent of development work, up from less than one percent in February, while more than 90 percent reached at least AL3. However, AL4 means that AI leads a task while humans still set its direction and approve the result; Claude reached AL5, the fully autonomous level, nowhere. The classifications were largely produced by Claude itself using evidence from internal communications and documents, and human assessments often differed from the model's judgment. The 26 percent figure is weighted by the amount of employee time spent on tasks, so it does not indicate how many decisions Claude makes or whether it determines research strategy. Anthropic also reports monitoring roughly 30,000 simultaneously running agents, with automated systems blocking 0.002 percent of more than a billion decisions in August, though the monitoring is still new and its coverage is uncertain. About six percent of AI research compute went to safety work in July, but Anthropic notes that this measure excludes some overlapping work and does not capture safety efforts that mainly require researcher time.
decoder