Overview

• A recent Frontiers in Communication study found reasoning-enabled large language models generate up to 50 times more CO₂ per query than concise-response counterparts.
• Top-performing systems like Cogito achieved higher accuracy but emitted three times more carbon than similarly sized models optimized for brevity.
• Emission levels varied by question type, with abstract or mathematical prompts producing as much as six times more CO₂ than simpler topics.
• Researchers advise users to choose task-specific smaller models and explicitly limit response length to reduce AI’s environmental impact.
• Experts warn that without formal transparency requirements, especially in the United States, per-prompt emissions data will remain largely unavailable.