Practical, data-backed steps to cut your AI energy use \u2014 without giving up the tools you rely on.
GPT-4o Mini handles most everyday tasks — summarisation, drafting, translation, Q&A — at 3× less energy than GPT-4o. Gemini Flash uses even less. You don't need a flagship model to write an email.
Reasoning models like o3, o4-mini, and DeepSeek-R1 “think” through chains of internal reasoning before answering. This is powerful for maths, logic, and complex analysis — but wildly wasteful for simple questions.
Every token in your prompt gets processed by billions of parameters. A 100-token prompt uses ~0.4 Wh. A 100,000-token prompt uses ~40 Wh. Long conversation threads compound this — each message re-processes the entire history.
Not all infrastructure is equal. Google runs Gemini on custom TPUs with 64% renewable energy matching — a median query uses just 0.24 Wh. Provider choice affects energy, carbon intensity, and water use.
Each API call or chat exchange has fixed overhead: network, GPU spin-up, context loading. One detailed prompt that covers everything is far more efficient than five iterative back-and-forth messages.
AI is transformative for creative work, analysis, and complex reasoning. But for simple factual lookups — “What's the capital of France?” — a search engine is faster, cheaper, and uses less energy.
Electricity grids are cleaner at certain times of day — typically when solar and wind generation peak. If you're running large batch jobs or fine-tuning, scheduling for low-carbon windows can meaningfully reduce your carbon footprint.
A cheat sheet you can come back to.
Calculate the energy, carbon, and water cost of your AI usage across 40+ models.