AI Footprint Estimator
A rough sense of the cost, electricity and water behind your AI prompts — with things you can actually picture.
This model's real cost is genuinely ambiguous — DeepSeek charges different peak and off-peak rates. The figure shown is the peak rate; off-peak is roughly half.
Enter how many prompts to see an estimate.
Rough, illustrative figures only — see the FAQ. Nothing you enter leaves your browser.
A very short choose-your-own-disaster. Someone is wrong on the internet. What happens next is up to you.
—
AI Footprint Estimator — Frequently Asked Questions
How accurate are these estimates?
They're rough and illustrative, not measurements. Cost uses each model's published API list price — input and output tokens priced separately, since they differ a lot — for a prompt of roughly 500 input + 300 output tokens. List prices were checked against public pricing trackers in August 2026 and change constantly: providers cut prices often, and some (DeepSeek especially) charge different peak and off-peak rates. Electricity uses about 0.3 Wh for a short standard prompt (Google put a median Gemini text prompt at ~0.24 Wh in 2025; OpenAI around 0.34 Wh), scaled up for large or heavy reasoning models and down for flash, lite and nano tiers. Water tracks the energy figure at about 1 ml per Wh, matching Josh-Fun's AI Water Converter. Real figures vary enormously by data centre, region, grid mix, prompt length and time of day.
Why is water involved?
Large data centres use water for evaporative cooling, and thermal electricity generation is water-intensive too. The number here is a simple per-prompt approximation for a sense of scale, not a facility measurement.
Which models are included?
The line-up as of August 2026 — OpenAI's GPT-5.6 Sol, Terra and Luna, plus GPT-5 and GPT-5 nano; Anthropic's Claude Opus 5, Sonnet 5 and Haiku 4.5; Google's Gemini 3 Pro, Gemini 3.7 Flash and 3.1 Flash-Lite; xAI's Grok 4.6; DeepSeek V4 Pro and V4 Flash; Zhipu's GLM-5.3; Moonshot's Kimi K3; Alibaba's Qwen3.8 Max; NVIDIA's Nemotron 3 Ultra; a hosted Llama 3.3 70B and Mistral Large 3; and a rough figure for generating one AI image. Bigger models and heavier reasoning runs cost more and use more energy per prompt — that's reflected. DeepSeek V4 Pro is flagged "price varies" because its peak and off-peak rates differ a lot; the figure shown is the peak rate.
What is the Dilemma Simulator?
A short comedic decision tree — "The Internet's Worst Decision". Make a few choices about how to handle someone being wrong online and you get an internet archetype at the end, ready to share.