The Hidden Cost of the Prompt

The largest models are remarkable, but they're not always necessary. Most routine tasks such as summarising text, checking grammar, answering a factual question, can be handled by smaller, lighter models that use a fraction of the compute. Using a flagship model for a simple question is a bit like driving a lorry to buy a pint of milk. The smaller "flash" or "haiku" tier models exist precisely because not every problem is a hard one. For the majority of everyday tasks, they perform just as well.

Similarly, a dedicated translation tool or a basic spellchecker will almost always be more efficient than routing the same job through a general-purpose LLM.

Every regeneration doubles the energy cost of that answer. Vague prompts tend to produce vague responses, which tend to produce another attempt, then another. Being precise the first time by describing the context, the format, the tone, the constraints isn't just good practice for getting a useful output. It's also the single most effective thing most users can do to reduce unnecessary computation. Tell the model how to think through the problem. The fewer times you need to ask, the lower the footprint.

Text generation and image generation are not comparable in energy terms. Generating one image uses roughly 100 times the energy of a text response. Video is higher still, a six-second AI clip can consume as much power as 600 standard searches. This isn't an argument against using these tools, but it's worth being honest about what "I'll just generate a few options" costs. When a stock photo or a simple graphic serves the purpose, it serves the planet better too.

AI-assisted search is genuinely useful for synthesis by pulling together complex, multi-source answers that a traditional search engine handles poorly. But for simple lookups, a traditional search is substantially greener. A weather forecast, a sports result, an address, these don't need the overhead of a language model. If your search interface automatically generates AI summaries for every query, most browsers and engines now let you disable that behaviour in settings. It's a small change with a consistent effect.

Those running batch jobs or large local models have a further lever: timing. Electricity grids vary in their carbon intensity across the day, solar-heavy grids are cleanest around midday, wind-heavy ones often cleaner overnight. Running intensive tasks during those windows means the same computation has a lower carbon cost. Running smaller, open-weight models locally eliminates the data centre overhead, the cooling, networking, and shared infrastructure entirely.