Thinking on a Budget: Revealed Preferences over Computational Resources in Large Language Models

Delegating work to an AI agent also delegates decisions about how it works: what to look up, how long to think, and how much to double-check. We study the economic rationality of these decisions and whether they are aligned with what actually improves task performance. In an augmented revealed-preference experiment, five open-weight language models repeatedly choose how to spend a fixed budget on worked examples, deliberation, and verification as the costs of these resources vary, and then solve reasoning tasks with the resources they chose. A separate calibration measures the performance of every possible allocation, so the same choices can be judged both for consistency and for effectiveness. We find that resource choices are far more coherent than matched random choice, and that every model devotes the largest share of its budget to deliberation, the one resource that reliably improves performance. Yet coherent demand is not always well targeted: some models devote much of their budget to resources that do little for performance, others systematically depart from higher-performing resource mixes, and all five buy a second round of verification even though it lowers their accuracy. Our results show that AI resource allocation can exhibit meaningful economic structure without being well aligned with task performance.