Skip to main content
Frugal AI Calculator

How to reduce your AI footprint

Your individual AI usage is a fraction of corporate consumption, but establishing efficient habits now scales well as these tools become unavoidable. Here is how to cut your footprint without losing productivity.

1. Match the model to the task

Do not use a massive frontier model (like OpenAI o1 or Gemini Argon) just to fix a typo or draft a standard email. Default to smaller, faster models like Claude Haiku or Gemini Flash for everyday tasks to use a fraction of the energy.

Take this further by exploring Small Language Models (SLMs). SLMs specialise in doing one or two things exceptionally well and operate far more efficiently than general-purpose LLMs.

2. Reserve reasoning modes for complex logic

Advanced reasoning models evaluate multiple paths before showing you an answer. This invisible compute tax is massive. Standard “thinking” modes cost roughly 30x more energy than normal text generation, while full reasoning modes can cost up to 300x more. Turn them off for simple tasks.

3. Write greener prompts

Stop saying please and thank you. The LLM does not give a shit.

Instead of sending ten short messages in a conversational chain, batch your context and instructions into a single, well-structured prompt. Fewer inference requests mean less server overhead.

You can use the CO-STARR framework to structure your requests. Research indicates that clear, formatted prompting reduces token use—and therefore energy—by up to 48%:

  • Context: Provide crisp background info in a single pass. Skip sweeping historical backstories that bloat the input token count.
  • Objective: State the exact goal. Ambiguity forces the AI to guess, triggering heavy follow-up correction loops.
  • Style & Tone: Define the register upfront. Request direct language and tell the model to drop conversational filler.
  • Audience: Naming the reader prevents the model from over-explaining basic concepts, naturally capping response length.
  • Response Format: Explicitly mandate the layout (e.g., bullet points or a specific JSON schema). Specifying format slashes wasted generation tokens by up to 45%.
  • Readability constraint: Add a firm guardrail like “Be direct; do not elaborate beyond what is asked” to prevent runaway generation.

4. Generate media sparingly

Generating a single high-resolution AI image uses as much energy as fully charging your smartphone. It is roughly 30x more energy-intensive than producing text. Video is astronomically higher: an 8-second video costs 1,000x more compute than a typical text prompt. Only generate media when absolutely necessary, and avoid spamming the regenerate button.

5. Further reading