AI Model Sustainability & CO2 Consumption

This document provides comprehensive data on the environmental impact of AI models used in the SF Explorer application, including CO2 emissions, water consumption, and energy usage.

Overview

AI model inference has a measurable environmental footprint. Understanding these impacts helps make informed decisions about model selection, balancing performance needs with sustainability goals.

Key Metrics

• CO2 Emissions (gCO2e): Grams of CO2 equivalent per 1,000 tokens
• Water Consumption (L): Liters of water per 1,000 tokens (cooling + energy production)
• Host Provider: Cloud infrastructure hosting the model
• PUE (Power Usage Effectiveness): Data center efficiency (lower is better, 1.0 is ideal)
• CIF (Carbon Intensity Factor): gCO2e per Wh of electricity

Model CO2 Consumption Data

OpenAI Models (Azure)

Model	CO2 (g/1k tokens)	Water (L/1k tokens)	Host	Source
GPT-5	13.778	17.686	Azure	Artificial Analysis (gpt-5-2025-08-07 high, DGX H200/H100)
GPT-5 Mini	7.746	6.102	Azure	Artificial Analysis (gpt-5-mini-2025-08-07 high, DGX H200/H100)
GPT-4.1	0.555	0.437	Azure	Artificial Analysis (gpt-4.1-2025-04-14, DGX H200/H100)
GPT-4.1 Mini	0.588	0.464	Azure	Artificial Analysis (gpt-4.1-mini-2025-04-14, DGX H200/H100)
GPT-4o	1.165	0.878	Azure	Artificial Analysis (gpt-4o-2024-11-20, DGX H200/H100)
GPT-4o Mini	0.639	0.463	Azure	Artificial Analysis (gpt-4o-mini-2024-07-18, DGX A100)
O3	0.990	0.780	Azure	Artificial Analysis (o3-2025-04-16, DGX H200/H100)
O4 Mini	5.132	4.043	Azure	Artificial Analysis (o4-mini-2025-04-16 high, DGX H200/H100)

Anthropic Models (AWS Bedrock)

Model	CO2 (g/1k tokens)	Water (L/1k tokens)	Host	Source
Claude 4.5 Sonnet	1.200	0.630	AWS	Artificial Analysis (claude-sonnet-4-5-20250929, DGX H200/H100)
Claude 4 Sonnet	1.178	0.620	AWS	Artificial Analysis (claude-sonnet-4-20250514, DGX H200/H100)
Claude 4.5 Haiku	0.784	0.413	AWS	Artificial Analysis (claude-haiku-4-5-20251001, DGX H200/H100)
Claude 3 Haiku	0.644	0.339	AWS	Artificial Analysis (claude-3-haiku-20240307, DGX H200/H100)

Google Models (Vertex AI)

Model	CO2 (g/1k tokens)	Water (L/1k tokens)	Host	Source
Gemini 2.5 Pro	1.543	1.838	Google	Artificial Analysis (gemini-2.5-pro, TPU V6e)
Gemini 2.5 Flash	0.556	0.663	Google	Artificial Analysis (google/gemini-2.5-flash, TPU V6e)
Gemini 2.5 Flash Lite	0.150	0.180	Google	Estimated based on Gemini 2.5 Flash
Gemini 2.0 Flash	0.450	0.540	Google	Estimated based on Gemini 2.5 Flash
Gemini 2.0 Flash Lite	0.120	0.145	Google	Estimated based on Gemini 2.5 Flash Lite

Amazon Models (AWS Bedrock)

Model	CO2 (g/1k tokens)	Water (L/1k tokens)	Host	Source
Amazon Nova Pro	0.500	0.350	AWS	Estimated based on similar models
Amazon Nova Lite	0.100	0.070	AWS	Estimated based on similar lightweight models

Environmental Factors by Cloud Provider

Different cloud providers have varying environmental impacts based on their infrastructure and energy sources:

Provider	PUE	CIF (gCO2e/Wh)	WUE (Site)	WUE (Source)	Notes
Azure	1.12	0.34	0.3	4.35	Higher carbon intensity
AWS	1.14	0.30	0.18	5.12	Moderate carbon intensity
Google	1.09	0.231	0.3	1.1	Lowest carbon due to renewable energy
xAI	1.5	0.385	0.36	3.142	Higher PUE
DeepSeek	1.27	0.6	1.2	6.016	Highest carbon intensity

Key Takeaways

• Google has the lowest carbon footprint due to extensive renewable energy investments
• Azure and AWS have comparable environmental profiles
• PUE measures data center efficiency (1.0 = perfect, typical range 1.1-1.5)
• CIF varies significantly based on regional energy grid composition

Sustainability Ratings

Models in the SF Explorer application are assigned sustainability ratings based on their CO2 emissions relative to other available models:

Rating	Percentile	Description
A+	Top 20%	Most sustainable - lowest emissions
A	21-40%	Very sustainable
B	41-60%	Moderate impact
C	61-80%	Higher impact
D	Bottom 20%	Least sustainable - highest emissions

Choosing Sustainable Models

Best for Sustainability (A+ rated)

1. Amazon Nova Lite - 0.100 g CO2/1k tokens
2. Gemini 2.0 Flash Lite - 0.120 g CO2/1k tokens
3. Gemini 2.5 Flash Lite - 0.150 g CO2/1k tokens
4. Gemini 2.0 Flash - 0.450 g CO2/1k tokens
5. Gemini 2.5 Flash - 0.556 g CO2/1k tokens

Best Balance of Performance & Sustainability

1. GPT-4.1 - 0.555 g CO2/1k tokens (excellent reasoning)
2. GPT-4.1 Mini - 0.588 g CO2/1k tokens (fast, efficient)
3. Claude 3 Haiku - 0.644 g CO2/1k tokens (quick responses)
4. GPT-4o Mini - 0.639 g CO2/1k tokens (multimodal)
5. Claude 4.5 Haiku - 0.784 g CO2/1k tokens (fast Claude)

High-Performance Models (Higher Emissions)

1. GPT-5 - 13.778 g CO2/1k tokens (flagship reasoning)
2. GPT-5 Mini - 7.746 g CO2/1k tokens (GPT-5 quality, faster)
3. O4 Mini - 5.132 g CO2/1k tokens (reasoning model)
4. Gemini 2.5 Pro - 1.543 g CO2/1k tokens (Google flagship)

Real-World Equivalents

To help contextualize emissions, here are equivalents for processing 1 billion prompts:

Model	Energy (MWh)	CO2 (Tons)	Water (kL)	Equivalent
GPT-5 (high)	12,156	4,133	56,136	~898 cars/year
GPT-4o Mini	564	192	1,910	~42 cars/year
Claude 3 Haiku	644	193	3,398	~42 cars/year
Gemini 2.5 Flash	723	167	995	~36 cars/year

Methodology

Data Sources

• Primary: Artificial Analysis benchmark data
• Hardware profiles: DGX H200/H100, DGX A100, TPU V6e
• Query length: Short (300 tokens), Medium (1000 tokens), Long (1500 tokens)

Calculation Formula

CO2 (gCO2e) = Energy (Wh) × CIF (gCO2e/Wh) × PUE
Water (mL) = Energy (Wh) × WUE × PUE

Assumptions

1. Query length normalization: Values normalized to per-1,000-token basis
2. Combined metrics: Uses Mean Combined Energy/Carbon/Water values (average of min/max utilization)
3. Estimated models: Some models without direct benchmarks use extrapolated values from similar architectures

References

• Artificial Analysis AI Benchmarks
• Google Sustainability Report
• Microsoft Azure Sustainability
• AWS Sustainability

---

Last updated: January 2026

Data source: packages/sf-explorer-app/src/Framework/utils/co2Consumption.csv