Anthropic has released new metrics regarding its AI development, revealing that Claude currently leads 26 percent of research tasks. This figure is based on an internal scale developed by Epoch AI, ranging from AL0 to AL5. As of August 2026, 26 percent of work reached the AL4 level, a significant increase from less than one percent in February 2026.
The AL4 level, described as AI leading, involves Claude performing tasks like bug fixing and testing without human intervention, though humans still provide direction and final approval. The scoring was conducted by Claude itself, which analyzed internal documents and Slack logs. Internal cross-checks showed that the model matched human assessments in 59 percent of cases.
The 26 percent figure represents the share of work measured by human hours spent on tasks. Anthropic also reported that its internal platform runs approximately 30,000 agents simultaneously. A real-time monitoring system blocked 0.002 percent of over one billion decisions made in August, while a secondary system flags about 100,000 logs per week for human review.
Regarding safety, Anthropic noted that six percent of compute resources for AI research were dedicated to safety work during a sample week in July. The company emphasizes that this metric is intended to provide transparency as CEO Dario Amodei advocates for a coordinated slowdown in frontier AI development, noting that safety research often requires more human time than compute.
