AI-optimized code used more energy than human code in 42 of 44 setups

A study accepted at the IEEE/ACM International Conference on Software Engineering (ICSE 2027) and posted to arXiv measured hardware-level energy as 11 open-source language models, each under four decoding settings, repeatedly rewrote code for speed on all 1,000 problems of the EffiBench benchmark. In 42 of the 44 model-and-setting configurations, the resulting code used more energy per problem than the human-written reference solution, with overheads from 4% to 886%; the two exceptions would need on the order of a million to ten million runs to pay back the energy spent generating them. Capping the number of output tokens cut inference energy by up to 69% against the default setting, and mixture-of-experts models used substantially less inference energy than dense models of similar size, the authors report.