The Efficient Frontier of AI Models: How to Select Models Using Portfolio Theory

Borrowing from Modern Portfolio Theory, the author constructs a Pareto efficient frontier for AI models, filtering out dominated models by plotting intelligence scores against comprehensive per-task costs. The analysis indicates that when evaluating token costs alone, many models appear uneconomical. However, factoring in human wait times—such as blocked, high-earning developers—significantly boosts the overall cost-performance of faster and higher-tier models. Furthermore, to address the high variance in long-context agent session costs, the author suggests using robust optimization to replace typical averages with worst-case costs, preventing budget overruns.