Overview

• A June 2025 study shows reasoning-enabled large language models generate up to 50 times more CO2 than concise response models by producing extensive ‘thinking’ tokens.
• Researchers measured that reasoning models averaged 543.5 tokens per query compared with 37.7 tokens for concise models, directly linking token count to carbon emissions.
• High-accuracy models like the Cogito 70B reached 84.9% correctness but emitted three times more CO2 than similarly sized concise models, highlighting an accuracy-sustainability trade-off.
• Complex subjects such as abstract algebra and philosophy drove emissions up to six times higher than straightforward topics like high school history.
• Users can reduce their carbon footprint by crafting concise prompts and selecting efficient models, and developers are urged to improve model efficiency and provide transparent emissions data.