Anthropic Starts Measuring How Much AI Does Inside AI Research

Anthropic has introduced an internal metric for estimating how much of its AI research and development work is performed by Claude. The company says the model “leads” 26% of its AI R&D work, based on a prototype index intended to track how AI contributes to building future AI systems.

The company also reported that, in a sample week in July, about 6% of its AI research compute went to safety work. For work carried out by AI itself, it said the share was 12%. Anthropic describes these figures as early measurements and says its aim is to give the public more visibility into changes inside frontier labs.

The measurement needs a denominator

The 26% figure is not a measure of autonomous research or of the share of a future model designed without people. Anthropic defines “leads” as cases where Claude can complete most of a task from a high-level prompt, while people remain involved in selecting tasks, reviewing outputs, and directing the process.

The company’s own numbers are therefore an initial reporting framework, not an independently audited industry standard. Comparisons across labs will be difficult unless they use consistent definitions for what counts as a task, completion, supervision, or meaningful research contribution.

Still, measuring internal AI use addresses a real transparency gap. If models increasingly help with coding, experiments, and analysis, outside observers need ways to track that change and its relationship to safety work. The next test is whether labs publish repeatable metrics that allow independent scrutiny over time.

Sources: Anthropic’s proposed R&D Automation Index; Reuters on the new metrics.