“For enterprises using large language models, the unit cost per token is decreasing, but aggregate spending is increasing due to more complex and longer-running queries.”