Aethir(@AethirCloud):A measurement study published in March put a number on something practitioners have felt for a while. Chain-of-thought prompting lifted accuracy sharply on hard tasks, and raised energy per answer by 120 to 150 times. Majority voting added 72% to 177% for much smaller gains. Same deployment, same question, two orders of magnitude apart in compute consumed. Any capacity plan that assumes a fixed cost per query is now wrong by default, sometimes by a hundredfold. The variance is in how the model thinks.

2026.10.31 11:51

A measurement study published in March put a number on something practitioners have felt for a while. Chain-of-thought prompting lifted accuracy sharply on hard tasks, and raised energy per answer by 120 to 150 times. Majority voting added 72% to 177% for much smaller gains. Same deployment, same question, two orders of magnitude apart in compute consumed. Any capacity plan that assumes a fixed cost per query is now wrong by default, sometimes by a hundredfold. The variance is in how the model thinks.