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.

Grid:Global avg · 400 gCO₂/kWh United States · 370 gCO₂/kWh EU · 250 gCO₂/kWh Low-carbon grid · 40 gCO₂/kWh

Assumptions: reference accelerator Cerebras Cerebras WSE-3 Turbo (43,200 GB/s HBM, 23000 W) · memory-bound decode at 50% MBU · batch 32 · 2-byte weights (FP16) · PUE 1.2 · grid 400 gCO₂/kWh (Global avg).

Model	Vendor	Active params	kWh / Mtok	gCO₂ / Mtok	$ / Mtok
Gemma 3 1B	Google	1B	0.018	8.9	—
Llama 3.2 1B	Meta	1B	0.018	8.9	—
Gemma 2 2B	Google	2B	0.037	17.7	—
Qwen3 30B A3B	Alibaba	3B	0.055	26.6	$0.50
Llama 3.2 3B	Meta	3B	0.055	26.6	—
Qwen3.6 35B A3B	Alibaba	3B	0.055	26.6	$1
Ministral 3 3B	Mistral	3B	0.055	26.6	$0.10
Qwen2.5 3B	Alibaba	3B	0.055	26.6	—
Gemma 3 4B	Google	4B	0.074	35.5	—
Gemma 4 26B A4B	Google	4B	0.074	35.5	$0.33
Mistral Small 4	Mistral	6B	0.111	53.2	$0.60
Mistral 7B	Mistral	7B	0.135	64.8	—
LLaMA 3 8B	Meta	8B	0.148	71.0	—
Ministral 3 8B	Mistral	8B	0.148	71.0	$0.15
Gemma 3 12B	Google	12B	0.222	106	—
Gemma 4 12B Unified	Google	12B	0.222	106	—
DeepSeek V4 Flash	DeepSeek	13B	0.240	115	$0.60
Mixtral 8x7B	Mistral	13B	0.240	115	$0.70
Ministral 3 14B	Mistral	14B	0.259	124	$0.20
LLaMA 4 Maverick	Meta	17B	0.314	151	$0.65
LLaMA 4 Scout	Meta	17B	0.314	151	$0.30
o4-mini	OpenAI	20B	0.370	177	$4.40
Qwen 3 235B (MoE)	Alibaba	22B	0.407	195	$1.82
Command A+	Cohere	25B	0.462	222	—
Gemma 3 27B	Google	27B	0.499	240	—
Gemma 4 31B	Google	31B	0.573	275	$0.35
Kimi K2.5	Moonshot	32B	0.592	284	—
Kimi K2.6	Moonshot	32B	0.592	284	$4
Qwen3 32B	Alibaba	32B	0.592	284	$0.28
Kimi K3	Moonshot	40B	0.739	355	$15
Qwen3.7 Max	Alibaba	40B	0.739	355	$3.75
Mistral Large 3	Mistral	41B	0.758	364	$1.50
Mixtral 8x22B	Mistral	44B	0.813	390	$6
DeepSeek V4 Pro	DeepSeek	49B	0.906	435	$1.98
GPT-4o	OpenAI	50B	0.924	444	$10
o3	OpenAI	50B	0.924	444	$8
LLaMA 3 70B	Meta	70B	1.29	621	—
Claude Haiku 3.5	Anthropic	70B	1.29	621	$4
Qwen 2.5 72B	Alibaba	72B	1.33	639	$0.40
Command R+	Cohere	104B	1.92	923	$10
Command A	Cohere	111B	2.05	985	$10
Mistral Large 2	Mistral	123B	2.27	1,091	—
Claude Opus 4	Anthropic	350B	6.47	3,106	$75
LLaMA 3.1 405B	Meta	405B	7.49	3,594	—

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.

Sizing report →Model catalog →

Modeled estimate; methodology above. Not investment or sustainability-reporting advice. Methodology.