Claude Leads 26% of Anthropic's Research Work, Focus on AI Safety

Claude, Anthropic's AI model, currently leads 26% of the company's research efforts, as quantified by the human hours involved. Dario Amodei, CEO of Anthropic, clarified that this leadership metric does not imply direct oversight but rather a significant contribution to future models. In February 2026, less than 1% of Anthropic's work was classified at AL4, but by August 2026, this figure rose to 26%. Furthermore, 90% of the work is reaching at least AL3, indicating a robust progression in AI capabilities.

In terms of safety measures, only 0.002% of decisions were halted by a real-time monitor in August, while a secondary monitoring system flagged 100,000 logs weekly, with 50 of those reaching human review. Notably, 6% of compute resources for AI research were allocated to safety work in July 2026. Anthropic operates 30,000 agents on its internal platform, enhancing its research and safety protocols. This emphasis on safety aligns with the company's stance that the burden of proof should rest with developers, a principle that is becoming increasingly relevant in AI governance discussions.