Claude leads 26% of Anthropic’s AI R&D, the company says
Claude’s role inside Anthropic has grown from helping researchers to leading substantial parts of their work, while humans remain in charge.
AI is moving deeper into the research process
Claude led less than 1% of Anthropic’s AI research and development work in February. By August 2026, that share reached 26%, the company says—a substantial shift in six months.[1]
“Leads” means AI handles most of a task under human supervision. Over 90% of work involved at least substantial AI collaboration; no measured category was fully autonomous.[1]
The distinction is about who carries a task forward. A researcher may still choose the question and judge the result even when an AI handles much of the work needed to get there. Counting AI use alone would hide that difference.
Measuring a job made of many different tasks
The approach draws on work by Epoch AI researchers, who proposed breaking AI research into more than sixty tasks across six categories. Their argument was practical: a benchmark cannot capture everything involved in coordinating research, running experiments and keeping the underlying systems working.[3]
That makes the unit of measurement important. Writing code for an experiment, monitoring a training run and deciding what to investigate next are different responsibilities. Epoch’s task-based approach is intended to show which parts of the research process AI can handle, instead of treating the whole profession as one activity.[3]
Anthropic weights categories by estimated human time. Its figure measures automation, not staffing cuts or hours saved. Using its own models as judges also risks missing mistakes those models share.[1]
The feedback loop inside an AI lab
If better models make research faster, that research could help produce better models in turn. The prospect matters because development might accelerate without a matching increase in the number of people able to review every decision.
People still supervise the work. Anthropic plans to give independent evaluators access to internal systems and data to check its practices and measurements.[1]
For readers outside the lab, the value of the index is a more concrete view of that changing relationship: how much work researchers hand over, and how much responsibility they retain.
Sources & context
Based on the sources below, with LLMCurrent’s analysis. Original reporting dated .
- Anthropic · research and methodology (opens in a new tab)
- National Assembly Library · publication record (opens in a new tab)
- Epoch AI researchers · measuring research automation (opens in a new tab)
About this article
Prepared with AI assistance from the cited material. It does not claim firsthand reporting or independent testing. Source claims are attributed; interpretation is identified in the text. Contact us about a correction.