“A Databricks paper argues that "cost per completed task" is a more useful metric for evaluating LLMs than "cost per token."”