How often does Claude use tools?

SemiAnalysis examined about 2.27 million internally collected responses from Anthropic’s Claude models. Its post and chart report that average client tool calls per response fell from 1.14 for Opus 4.6 to 0.76 for Opus 4.8, while Fable 5 averaged 1.00. An average of one does not mean every response used a tool: some responses can contain several calls and others none. The chart separately shows the share of responses with at least one tool call falling from 78.8% to 55.5% between Opus 4.6 and Opus 4.8. These figures describe that internal usage sample; without matched tasks, they cannot establish whether a model is more capable or efficient.