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 | 0.018 | 0.9 | — | |
| Llama 3.2 1B | Meta | 1B | 0.018 | 0.9 | — |
| Gemma 2 2B | 2B | 0.037 | 1.8 | — | |
| Qwen3 30B A3B | Alibaba | 3B | 0.055 | 2.7 | $0.50 |
| Llama 3.2 3B | Meta | 3B | 0.055 | 2.7 | — |
| Qwen3.6 35B A3B | Alibaba | 3B | 0.055 | 2.7 | $1 |
| Ministral 3 3B | Mistral | 3B | 0.055 | 2.7 | $0.10 |
| Qwen2.5 3B | Alibaba | 3B | 0.055 | 2.7 | — |
| Gemma 3 4B | 4B | 0.074 | 3.5 | — | |
| Gemma 4 26B A4B | 4B | 0.074 | 3.5 | $0.33 | |
| Mistral Small 4 | Mistral | 6B | 0.111 | 5.3 | $0.60 |
| Mistral 7B | Mistral | 7B | 0.135 | 6.5 | — |
| LLaMA 3 8B | Meta | 8B | 0.148 | 7.1 | — |
| Ministral 3 8B | Mistral | 8B | 0.148 | 7.1 | $0.15 |
| Gemma 3 12B | 12B | 0.222 | 10.6 | — | |
| Gemma 4 12B Unified | 12B | 0.222 | 10.6 | — | |
| DeepSeek V4 Flash | DeepSeek | 13B | 0.240 | 11.5 | $0.60 |
| Mixtral 8x7B | Mistral | 13B | 0.240 | 11.5 | $0.70 |
| Ministral 3 14B | Mistral | 14B | 0.259 | 12.4 | $0.20 |
| LLaMA 4 Maverick | Meta | 17B | 0.314 | 15.1 | $0.65 |
| LLaMA 4 Scout | Meta | 17B | 0.314 | 15.1 | $0.30 |
| o4-mini | OpenAI | 20B | 0.370 | 17.7 | $4.40 |
| Qwen 3 235B (MoE) | Alibaba | 22B | 0.407 | 19.5 | $1.82 |
| Command A+ | Cohere | 25B | 0.462 | 22.2 | — |
| Gemma 3 27B | 27B | 0.499 | 24.0 | — | |
| Gemma 4 31B | 31B | 0.573 | 27.5 | $0.35 | |
| Kimi K2.5 | Moonshot | 32B | 0.592 | 28.4 | — |
| Kimi K2.6 | Moonshot | 32B | 0.592 | 28.4 | $4 |
| Qwen3 32B | Alibaba | 32B | 0.592 | 28.4 | $0.28 |
| Kimi K3 | Moonshot | 40B | 0.739 | 35.5 | $15 |
| Qwen3.7 Max | Alibaba | 40B | 0.739 | 35.5 | $3.75 |
| Mistral Large 3 | Mistral | 41B | 0.758 | 36.4 | $1.50 |
| Mixtral 8x22B | Mistral | 44B | 0.813 | 39.0 | $6 |
| DeepSeek V4 Pro | DeepSeek | 49B | 0.906 | 43.5 | $1.98 |
| GPT-4o | OpenAI | 50B | 0.924 | 44.4 | $10 |
| o3 | OpenAI | 50B | 0.924 | 44.4 | $8 |
| LLaMA 3 70B | Meta | 70B | 1.29 | 62.1 | — |
| Claude Haiku 3.5 | Anthropic | 70B | 1.29 | 62.1 | $4 |
| Qwen 2.5 72B | Alibaba | 72B | 1.33 | 63.9 | $0.40 |
| Command R+ | Cohere | 104B | 1.92 | 92.3 | $10 |
| Command A | Cohere | 111B | 2.05 | 98.5 | $10 |
| Mistral Large 2 | Mistral | 123B | 2.27 | 109 | — |
| Claude Opus 4 | Anthropic | 350B | 6.47 | 311 | $75 |
| LLaMA 3.1 405B | Meta | 405B | 7.49 | 359 | — |
Sorted most-efficient first. Only models with a disclosed active-parameter count are shown (44 of 72); 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.