How we calculate your AI footprint

The short version

For every message in your export, we look up a published estimate of how much energy, water, and CO₂ a single message of that model type costs. Then we add them all up and scale CO₂ to your local electricity grid. That’s it — no machine learning, no token counting, no calls back to OpenAI or Anthropic. Pure arithmetic over a small table of numbers.

The per-message estimates

Each message is assigned a “cost” based on which model answered it:

• Energy — from Epoch AI’s 2025 analysis (“How much energy does ChatGPT use?”), which put GPT-4o at ~0.3 Wh per query. Cross-checked against Sam Altman’s July 2025 figure of 0.34 Wh.
• Water — calculated as energy × 3.7 mL/Wh, using the Scope 1 + 2 method from Ren et al.’s peer-reviewed paper “Making AI Less Thirsty” (CACM 2025). This counts both on-site datacenter cooling and water consumed at the power plants that supply the electricity — the same accounting convention we use for CO₂.
• CO₂ — derived from the energy figure at a world-average grid carbon intensity, then multiplied by a regional factor (see below).
• Reasoning models (o1, o1-preview, o1-mini) — extrapolated at roughly 30–50× the GPT-4o cost, consistent with Jegham et al. (2025), who measured reasoning models at ~30× non-reasoning energy on long prompts.

Your local electricity grid matters for CO₂

Because generating electricity produces very different amounts of CO₂ depending on the energy mix, we multiply the raw CO₂ total by a regional factor derived from IEA grid data:

• US — ×0.95
• UK / Europe — ×0.44
• India — ×1.77
• Australia — ×1.38
• Other / unknown — ×1.0

So a UK user’s CO₂ number is less than half a US user’s, for the same number of queries. Energy and water are not region-adjusted — those numbers represent the actual resource use, not the emissions profile.

The everyday comparisons

To make the numbers less abstract, we convert them to familiar things:

• Bottles of water — 500 mL per bottle. Used when your total is ≥ 4 L.
• Glasses of water — 250 mL per glass. Used below 4 L.
• Phone charges — 10 Wh (0.01 kWh) per charge of a typical smartphone battery.
• Hours of Netflix — 77 Wh (0.077 kWh) per hour of HD streaming (IEA 2020 estimate). Shown when your energy is below one phone charge.
• Miles driven — 404 g CO₂ per mile. This is the US EPA average for a gasoline passenger vehicle (~22 MPG), from the EPA’s “Greenhouse Gas Emissions from a Typical Passenger Vehicle” factsheet.
• % of a daily footprint — 44 kg CO₂ per day is the average American’s total daily footprint across everything (commute, food, home energy). So 440 g would be 1% of a typical day.

What we don’t count (and why your number is still an estimate)

• Training energy. We count only the cost of running your queries. The energy cost of training GPT-4o or Claude is amortised across billions of users and excluded here — include it and the numbers roughly double.
• Message length. Every message is priced the same, whether it’s “hi” or a 3,000-word analysis. A token-weighted model would be more accurate but isn’t feasible with the export data we get.
• Image generation, voice, tools, browsing. These are typically more expensive than text but aren’t separable in the export.
• ChatGPT model detection. OpenAI’s new export format no longer includes the model slug per message. For those uploads, every message resolves to the GPT-4o default — even if you actually used GPT-4, o1, or o1-mini for some of them. Your reasoning-model usage, specifically, is almost certainly under-counted.
• Claude model detection. Claude exports include no model information at all — not even a blank field. Claude usage is all billed at the GPT-4o baseline. Sonnet, Opus, and Haiku all look identical to us. Opus is probably larger than GPT-4o, so heavy Opus users are likely under-counted too.

Sources

• Epoch AI (2025). How much energy does ChatGPT use?
• Ren, S. et al. (2023 / CACM 2025). Making AI Less Thirsty — peer-reviewed Scope 1+2 water methodology.
• Jegham, A. et al. (2025). How Hungry is AI? — empirical energy measurements including reasoning models.
• Altman, S. (2025). The Gentle Singularity — OpenAI’s own per-query figures.
• IEA (2024). Energy and AI — regional grid carbon intensity.
• US EPA. Greenhouse Gas Emissions from a Typical Passenger Vehicle — 404 g CO₂/mile figure.