Anthropic has published data on the use of Claude in the company’s internal AI research and development. According to the data, in 26% of the work, Claude can carry out the process largely end to end, starting from a high-level prompt; however, oversight remains with human employees. AI systems perform major parts of the work in more than 90% of the research with close human guidance. Anthropic states that Claude was not fully autonomous in any of the measured work. The data was prepared using the “AI-led AI R&D” method, by combining the company’s usage data with Epoch AI’s automation scale. This is the first of three methods intended to compare the pace of AI development. The method shows Claude’s levels of automation since August 2025; similar measurements can be verified by independent third parties.
Why it matters
These data make it necessary to assess the use of AI in research and development not solely through the lens of full independence, but by looking at how much of the task is delegated with human guidance. Claude’s role within the company can thus be understood less as a model in which human oversight has disappeared and more as a way of working in which different levels of automation coexist. For companies and employees evaluating AI-assisted research processes, this distinction makes it possible to consider speed measurements separately from oversight responsibilities. However, because the current picture was created using only the first of three methods, it leaves open questions about how the other measurement approaches would frame the same use and the extent to which the results of independent verification would support the findings.
Background
Anthropic is not a new name in the FikirPilot archive: over the past 90 days, we have published 24 articles mentioning the name; the most recent was dated 19 September 2026.