AI agents typically rely on large language models to plan and act over multiple steps, but the way models are evaluated may be misaligned with that use. A new preprint on arXiv (2610.19345) notes that LLMs are usually compared per call—single responses—rather than per completed task. That leaves an open question: which model should an agent use to complete tasks reliably?

Rather than defaulting to the largest available model, the paper frames model choice as a 'right-sizing' problem. If a smaller model can finish a multi-step task with acceptable reliability, it may be the greener option, reducing compute and energy use. The title asks 'When is smaller greener?' and the abstract suggests the answer depends on task-level outcomes, not just single-call quality.

The work is a single source, so there are no conflicting findings to compare. It points toward a shift in benchmarking for agentic AI: evaluate models on the whole task, not just the next token.