Match the AI model to the task type to cut costs without losing quality

The core idea comes from a framework proposed by Andrej Karpathy. It centers on a concept called verifiability: some tasks have a clear right or wrong answer — like math problems or code tests — while others are open-ended, like writing an essay or giving strategic advice. For verifiable tasks, you can use a cheaper model and simply retry if it gets it wrong. For non-verifiable tasks, you need a stronger model from the start because there's no easy way to catch a bad answer.

The experiment first classified each task into one of these two categories, then automatically sent it to the appropriate model. Across 120 sample tasks and 3 models, this routing approach was more cost-effective than always using the most powerful model. The practical takeaway for AI agent builders: assign each step of your pipeline to the cheapest model that can reliably handle it, and reserve expensive models only where judgment truly matters.