AI Carbon Footprint Calculator — All Models (2026)
CO2 emissions for GPT-4o, Claude, Gemini, Llama, image gen, video gen, and local models.
How AI carbon emission estimates are calculated
A query's carbon footprint depends on four things: model size (larger models require more GPU compute per token), inference hardware (A100 vs H100 vs TPU efficiency), data center PUE (power usage effectiveness, how much extra energy is spent on cooling and overhead), and the carbon intensity of the local electricity grid.
The numbers in this calculator are median estimates from published research — Patterson et al. (2021), Lannelongue et al. Green Algorithms, and subsequent industry studies. They assume average grid conditions. Google and Microsoft both claim their data centers run on matched renewable energy, which can reduce effective CO2 to near zero under their accounting methods — though how "renewable" is counted varies. DeepSeek operates efficient data centers in China with a grid mix that's improving but still partially coal-dependent.
Which AI tasks use the most energy
Reasoning models (o1, o3, DeepSeek R1) use significantly more compute per query than standard chat models — they run extended "thinking" chains before responding, sometimes for 30–120 seconds. An o1 query can use 3–8x the energy of a comparable GPT-4o query.
Image generation is more energy-intensive than text: diffusion models run many iterative denoising steps, each requiring GPU compute. A Midjourney or DALL-E 3 image is roughly 2–30x a text query depending on resolution and steps. Video generation is the most intensive by far — a short Sora-style clip can represent the energy cost of 50+ standard text queries.
At the other end: embedding models, classification tasks, and small models (Gemini Flash, GPT-4o mini, Haiku) use a tiny fraction of flagship model energy. Choosing the smallest capable model is the single highest-leverage environmental decision an AI developer can make.
Self-hosted vs cloud: who pays the carbon cost
When you use a cloud API, the provider's data center absorbs the energy load — potentially on cleaner power than your local grid. When you self-host on a gaming GPU or standard cloud VM, the carbon cost shifts to your infrastructure. If you're in a high-renewable region (Pacific Northwest, Scandinavia, parts of Western Europe), local inference can be very clean. On a coal-heavy grid, running models locally is often more carbon-intensive than a major hyperscaler's renewable-powered data center.
Putting AI emissions in real-world perspective
20 GPT-4o queries/day = roughly 22kg CO2/year from AI. A gas car driving 12,000 miles/year emits roughly 4,800kg. One transatlantic flight emits 1,000–2,000kg. At an individual level, AI usage is a small fraction of your total carbon footprint.
At the aggregate level, it's a different story. Hundreds of millions of daily AI users add up to a meaningful electricity draw, and AI data center electricity consumption is one of the fastest-growing industrial loads globally. The meaningful environmental decisions happen at the infrastructure level — renewable energy sourcing, hardware efficiency improvements, model distillation to reduce compute needs — not primarily at the individual query level.
Training GPT-4 is estimated to have emitted ~500 tonnes CO2 — equivalent to 500 transatlantic flights. At scale, cumulative inference emissions overtake training emissions within months of deployment. Model efficiency (smaller models for appropriate tasks) reduces both cost and emissions simultaneously.
Emission estimates based on Patterson et al. (2021), Lannelongue et al. Green Algorithms framework, and subsequent industry analysis. Figures are midpoint estimates; actual emissions vary 3-5x depending on data center location, hardware generation, and grid carbon intensity.