Anthropic reports that its Claude models now lead 26% of internal research and development tasks, categorized as AL4 level automation. This figure marks a significant increase from less than 1% recorded in February 2026.
As of August 2026, over 90% of the company's R&D tasks involved AI collaboration at AL3 level or higher. During the same period, approximately 30,000 AI agents were active on the internal platform, though no fully autonomous AL5 work has been reported.
In April 2026, Claude Opus 4.6 agents addressed 97% of the performance gap in an alignment problem within seven days. Additionally, August 2026 tests showed Claude outperforming human experts in designing protein binders.
The company projects that Claude could lead approximately 80% of its internal AI R&D by the end of 2026. This shift would transition the role of human researchers toward review and decision-making processes.
