An internal tool classifies incoming messages into one of eight categories, at roughly two million messages a month. The team is currently using the largest available foundation model and the bill has become the main objection to the project. What is the most promising first change?

• AIncrease the context window so fewer calls are needed.
• BEvaluate a smaller model on the same task, since narrow classification often does not need frontier capability.
• CRaise temperature so the model reaches an answer in fewer tokens.
• DMove to provisioned throughput without changing the model.