Anthropic's AI R&D Metrics

As of August 2026, Claude is responsible for 26% of AI research and development work at Anthropic. The company reports that over 90% of its tasks are at or above the level of 'AI collaborates', indicating a high degree of AI involvement in its operations.

Anthropic employs approximately 30,000 agents engaged in research and engineering tasks, all of which are monitored online, ensuring 100% coverage of agent actions before execution. The monitoring systems have blocked only 0.002% of decisions, reflecting a highly permissive operational environment.

Human reviews of blocked actions are conducted within one week, demonstrating a commitment to rapid oversight. During the period from July 13 to July 20, 2026, 6% of compute resources were allocated to safety work, while 12% of compute dedicated to AI-driven R&D was also focused on safety measures.

Dario Amodei, CEO of Anthropic, stated plans to integrate independent third-party evaluators from multiple organizations into their processes. Despite the extensive use of AI, Anthropic clarified that as of August 2026, Claude is not operating fully autonomously for any subset of AI R&D work. This follows previous discussions on AI safety and oversight, emphasizing the need for rigorous evaluation in AI development.