Anthropic released figures showing how directly its AI model is being deployed in AI research and development. Claude is effectively leading a substantial share of R&D work, with tens of thousands of AI agents simultaneously conducting research and engineering inside the company.

Online outlet Gigazine reported on Thursday that Anthropic said 26 percent of its R&D work had reached level 4 on its "Anthropic R&D Automation Index" as of August 2026. At that level, humans set broad direction and AI performs most of the work on its own.

Anthropic classifies AI participation in R&D on a scale of 0 to 5. Level 0 is where humans perform all tasks, while level 5 is full autonomy, with AI handling everything from identifying tasks to execution.

Anthropic said there was almost no level 4 work until earlier this year, but it rose to 26 percent in August. It added that levels 3 and above, where humans and AI collaborate, accounted for more than 90 percent of total R&D work. It said there were no cases yet that had reached level 5, the fully autonomous stage.

The scale of internal AI agents also stands out. On Anthropic's main internal agent platform, about 30,000 AI agents that perform research and engineering tasks are running simultaneously.

Anthropic said it also operates monitoring systems for agents as AI use expands. It combines checks before execution with analysis of activity records after execution. In August 2026 alone, it monitored more than 1 billion decisions made by AI agents, and humans directly reviewed about 50 high-priority cases each week.

Anthropic also disclosed the share of computing resources for AI R&D that are allocated to safety research. Based on its tally for the week from July 13 to 20, about 6 percent of total AI R&D compute was used for safety-related research. Looking only at work in which AI conducts AI R&D, the share allocated to safety research was about 12 percent.

The disclosure shows AI moving beyond a simple development support tool and becoming deeply embedded in the R&D process inside AI companies. It also indicates that the pace of AI research automation is rising quickly, with level 4 automated work expanding to 26 percent in just a few months from almost none at the start of the year.

Anthropic drew a line, saying it has not yet reached fully autonomous R&D. It also stressed that it is building control systems for expanded AI use by running agent monitoring and safety research in parallel.

Anthropic said it plans to continue disclosing related metrics and to accept independent third-party evaluations. It said the focus will be not only how much AI's share of R&D rises further, but also how the scale of agent monitoring and compute allocated to safety research change.