Preprint finds longer agent tasks can reverse smaller models' energy advantage

A preprint testing 15 open-weight language models found that task length can reverse which model uses less energy per successfully completed agent task. Its framework accounts for failed attempts and predicts when differences in reliability outweigh savings on individual model calls. In two synthetic tasks with measured GPU energy, predicted crossover points fell within the observed range for 72 of 88 crossovers. At 20 steps, a 3-billion-parameter model used a tenth of the energy of a 27-billion model per completed task. At 100 steps, a 1.7-billion model used almost four times as much as an 8-billion model.