Anthropic shared the latest advancements in AI automation within Claude’s research and development (R&D) process in a blog post. The company has established an automation level classification system spanning from AL0 to AL5 to gauge progress. As of August 2026, Claude has taken the lead in 26% of its AI R&D tasks, with over 90% of all R&D work having at least reached the stage of ‘AI collaboration.’ At present, no work has achieved the fully autonomous AL5 level. In February 2026, less than 1% of R&D tasks had entered the AI-led AL4 phase, but this figure surged to 26% within just half a year. On Anthropic’s primary internal agent platform, roughly 30,000 AI agents are actively engaged in R&D work simultaneously at any given moment, collectively making over 1 billion decisions in August alone. These agents are capable of collaborating and delegating tasks. All actions are subjected to online monitoring before execution, with only 0.002% of actions being halted. Additionally, an offline monitoring mechanism is in place, generating approximately 50 high-priority records each week for manual review. When it comes to computational resource allocation, about 6% of the computational power dedicated to AI R&D is allocated to safety research. In AI-driven AI R&D, this proportion rises to approximately 12%. Currently, humans retain the ultimate decision-making authority in R&D, but the division of labor between humans and AI is undergoing rapid transformation.