XNews

Anthropic internal AI development metrics dashboard

Anthropic internal AI development metrics dashboard

Anthropic disclosed its internal dashboard, showing that as of July 2026, Claude leads 26% of AI R&D work (AL4), with over 90% of tasks involving AI at AL3 or above. The dashboard also tracks agent oversight, with 30,000 agents running concurrently and 6% of AI R&D compute allocated to safety.

Key insights

AI Dominates 26% of Anthropic's R&D Work, Nearing Autonomy

Anthropic's internal dashboard, first made public, reveals that as of July 2026, Claude leads 26% of the company's AI R&D work, achieving AL4 (AI-dominated) status. Over 90% of R&D tasks involve AI at AL3 or above, indicating deep collaboration or dominance. However, no work subset has reached AL5 (full autonomy). The metric is based on a weekly sample of 20% of R&D staff, totaling about 15,000 tasks, and is weighted by human time. This baseline provides a direct window into the pace of AI self-improvement.

Safety Monitoring: 30,000 Agents, 0.002% Intervention Rate

Anthropic's oversight system involves two layers: online monitoring intercepts 100% of agent actions before execution, analyzing over 1 billion decisions in August 2026, with an intervention rate of only 0.002% (about 1 in 47,000). Offline monitoring flags about 100,000 conversation logs weekly, with roughly 50 escalated to human review. The low intervention rate could indicate either a well-functioning system or overly lenient monitoring, a risk the report acknowledges. The system is only a few months old, so its effectiveness remains under evaluation.

Compute Allocation: 6% of AI R&D Spend Goes to Safety

Anthropic's dashboard includes a hard metric: of all compute tasks for AI R&D in a week (July 13-20, 2026), about 6% was allocated to safety work. When considering only AI-driven AI R&D, the safety share rises to 12%. These are conservative estimates. The compute allocation is considered the most verifiable and least gameable metric, enabling cross-lab comparisons and potential policy limits on AI R&D compute shares.

Source posts