Anthropic publishes internal AI development metrics dashboard

Anthropic released an internal dashboard showing AI now leads 26% of its R&D work, with over 90% of tasks involving AI at AL3 or above. The dashboard tracks R&D automation, agent oversight of 30,000 concurrent agents, and compute allocation with 6% for safety.
Key insights
R&D automation index shows AI dominance but not full autonomy
Anthropic's dashboard measures AI's role in its own development using the AL0-AL5 scale. As of July 2026, 26% of R&D tasks reached AL4 (AI-led with human supervision), and over 90% were at AL3 or above. No tasks reached AL5 (full autonomy). This baseline allows tracking acceleration of recursive self-improvement.
Agent oversight system monitors 30,000 concurrent agents
Anthropic's internal platform runs about 30,000 AI agents simultaneously for R&D and engineering. Online monitoring intercepts 100% of actions before execution, with a 0.002% interception rate (about 1 in 47,000). Offline auditing flags about 100,000 conversations weekly, with roughly 50 escalated to human review. The low interception rate could indicate either good system behavior or overly lenient monitoring.
Compute allocation reveals 6% for safety work
Anthropic analyzed one week of compute tasks in July 2026, finding about 6% of AI R&D compute went to safety work. For AI-driven AI R&D specifically, safety accounted for about 12%. The methodology provides a verifiable baseline for cross-lab comparisons and potential safety compute commitments.
Claude's role in R&D grew from 1% to 26% in six months
The dashboard shows Claude's share of model development tasks rose from 1% to 26% in just six months. Claude now participates in or leads over 90% of Anthropic's model development work. This rapid increase highlights the accelerating integration of AI into its own development pipeline.