Reference · Sustainability
The carbon twin of $/M-token
Every token has an energy shadow. This is a modeled estimate of the electricity (kWh) and emissions (gCO₂) per million output tokens for AI models with disclosed active-parameter counts, normalized on a reference accelerator. Like our sizing and teardowns: modeled, assumptions published — best for relative comparison, not a billed figure.
| Model | Vendor | Active params | kWh / Mtok | gCO₂ / Mtok | $ / Mtok |
|---|---|---|---|---|---|
| Gemma 3 1B | 1B | 0.004 | 1.2 | — | |
| Llama 3.2 1B | Meta | 1B | 0.004 | 1.2 | — |
| Gemma 2 2B | 2B | 0.008 | 2.3 | — | |
| Qwen 3 30B (MoE) | Alibaba | 3B | 0.012 | 3.5 | $0.50 |
| Qwen2.5 3B | Alibaba | 3B | 0.012 | 3.5 | — |
| Llama 3.2 3B | Meta | 3B | 0.012 | 3.5 | — |
| Gemma 3 4B | 4B | 0.016 | 4.7 | — | |
| Mistral 7B | Mistral | 7B | 0.029 | 8.6 | — |
| LLaMA 3 8B | Meta | 8B | 0.031 | 9.4 | — |
| Llama 3.1 8B | Meta | 8B | 0.031 | 9.4 | — |
| Gemma 3 12B | 12B | 0.047 | 14.1 | — | |
| LLaMA 4 Scout | Meta | 17B | 0.066 | 19.9 | $0.30 |
| LLaMA 4 Maverick | Meta | 17B | 0.066 | 19.9 | $0.80 |
| o4-mini | OpenAI | 20B | 0.078 | 23.4 | $4.40 |
| Qwen 3 235B (MoE) | Alibaba | 22B | 0.086 | 25.8 | $1.82 |
| Gemma 3 27B | 27B | 0.105 | 31.6 | $0.33 | |
| DeepSeek R1 | DeepSeek | 37B | 0.145 | 43.4 | $2.50 |
| DeepSeek V3 | DeepSeek | 37B | 0.145 | 43.4 | $0.80 |
| DeepSeek-V3 (MoE) | DeepSeek | 37B | 0.145 | 43.4 | — |
| Mixtral 8x22B (MoE) | Mistral | 39B | 0.152 | 45.7 | — |
| Mixtral 8x22B | Mistral | 39B | 0.152 | 45.7 | $6 |
| GPT-4o | OpenAI | 50B | 0.195 | 58.6 | $10 |
| o3 | OpenAI | 50B | 0.195 | 58.6 | $8 |
| Gemini 2.5 Flash | 60B | 0.234 | 70.3 | $2.50 | |
| LLaMA 3 70B | Meta | 70B | 0.273 | 82.0 | — |
| Claude Haiku 3.5 | Anthropic | 70B | 0.273 | 82.0 | $4 |
| Llama 3.1 70B | Meta | 70B | 0.273 | 82.0 | — |
| Qwen 2.5 72B | Alibaba | 72B | 0.281 | 84.4 | $0.40 |
| Grok-2 | xAI | 80B | 0.313 | 93.8 | $10 |
| Command R+ | Cohere | 104B | 0.406 | 122 | $10 |
| Claude Sonnet 4 | Anthropic | 120B | 0.469 | 141 | $15 |
| Mistral Large 2 | Mistral | 123B | 0.480 | 144 | $1.50 |
| Gemini 2.5 Pro | 300B | 1.17 | 352 | $10 | |
| GPT-4.5 | OpenAI | 300B | 1.17 | 352 | $15 |
| Grok-3 | xAI | 300B | 1.17 | 352 | $15 |
| Claude Opus 4 | Anthropic | 350B | 1.37 | 410 | $75 |
| LLaMA 3.1 405B | Meta | 405B | 1.58 | 475 | — |
| Llama 3.1 405B | Meta | 405B | 1.58 | 475 | — |
| Kimi K3 | Moonshot | 2800B | 10.94 | 3,281 | $15 |
Sorted most-efficient first. Only models with a disclosed active-parameter count are shown (39 of 46); most closed models don't publish parameters.
How it's modeled
Decode is memory-bandwidth-bound: each output token reads the model's active parameters from HBM once, amortized across the serving batch. So energy/token ≈ chip power ÷ (per-stream throughput × batch), where per-stream throughput = (HBM bandwidth × MBU) ÷ (active params × bytes/param). Emissions = energy × PUE × grid carbon intensity. Real serving efficiency varies widely with batch size, precision, hardware and grid — treat these as directional. Terms: MBU, PUE, grid carbon intensity, gCO₂ per token.
Model your workload's footprint precisely
For your actual model, chip, batch and region, the sizing report computes energy, cost and $/M-token together.
Modeled estimate; methodology above. Not investment or sustainability-reporting advice. Methodology.