Anthropic Publishes Internal AI Development Dashboard

Anthropic publicly disclosed its internal AI development dashboard for the first time, revealing that as of July 2026, Claude dominates 26% of AI R&D work (AL4 level), with over 90% of R&D involving AI deeply. At any moment, 30,000 agents work simultaneously. About 6% of compute for AI R&D is allocated to safety. The dashboard includes three metric groups: R&D automation index, agent oversight, and compute allocation.
Key insights
26% of AI R&D Work Reaches AL4 (AI-Dominated) Level
Anthropic's internal dashboard shows that 26% of AI R&D tasks are at AL4, meaning AI handles most work end-to-end with human supervision. Over 90% of R&D tasks are at AL3 or above (deep collaboration or dominance). No tasks reached AL5 (fully autonomous). The measurement was done by sampling 20% of R&D staff weekly, organizing 15,000 tasks into a 542-node tree, and rating them against Epoch AI's AL0-AL5 scale.
30,000 Agents Simultaneously Active with Two-Layer Oversight
At any given time, approximately 30,000 agents are performing R&D and engineering work on Anthropic's most-used platform. Online monitoring intercepts 100% of agent actions before execution, typically within seconds. In August 2026, over 1 billion decisions were analyzed with an interception rate of 0.002% (about 1 in 47,000), mostly for fast, irreversible harms like exfiltration of model weights. Offline monitoring flags about 100,000 conversations weekly, with ~50 per week escalated to human review.
Compute Allocation: 6% for Safety in AI R&D
For the week of July 13-20, 2026, Anthropic analyzed all ~10,000 compute tasks. About 6% of compute used for AI R&D was dedicated to safety work. When considering only AI-driven AI R&D, safety share was about 12%. These are conservative estimates. The methodology allows for cross-lab comparisons and potential future safety compute commitments.
Dashboard Provides Baseline for Measuring AI Self-Improvement Acceleration
The 26% AL4 figure is the most critical number, as it establishes a baseline for future comparisons. Anthropic notes that while AI participation in AI R&D is now normal, the critical threshold of recursive self-improvement (AL5) has not been reached. The dashboard's metrics are designed to be tracked over time to observe acceleration.
Low Interception Rate Raises Questions About Oversight Stringency
The 0.002% interception rate can be interpreted either as a well-functioning system or as overly lenient monitoring. Anthropic itself notes that at the scale of tens of thousands of agents, rare events occur frequently. The oversight system is only a few months old, and its effectiveness requires more time to evaluate.