Reducing AI Costs by Optimizing Cost per Task

OpenAI engineers explain how to evaluate AI costs by the expense of completing a task, rather than token prices alone. They cover model selection and application design choices that can affect cost while maintaining the quality a task requires.

■ Measuring cost and choosing models
- Cost per task, task success, and examples from customer applications
- Defining accuracy targets and comparing model configurations

■ Application-level cost controls
- Prompt caching and programmatic tool calling
- Reasoning effort, Batch and Flex APIs, and an iterative evaluation workflow

For developers and teams building AI applications, the discussion offers a framework for testing models and configuration changes. It does not provide a task-specific implementation; viewers can apply the suggested approach by defining an accuracy threshold, evaluating representative tasks, and changing one setting at a time.