AI Model Sustainability Guide

Understanding and minimizing the environmental impact of AI model usage.

---

Why Sustainability Matters

The Environmental Impact of AI

AI models, especially large language models (LLMs), require significant computational resources:

• Energy Consumption: Large models like GPT-5 can use 20-100x more energy than efficient models
• CO₂ Emissions: Training GPT-3 produced ~552 tons of CO₂; inference adds up at scale
• Water Usage: Data centers use water for cooling - up to 17 liters per 1,000 tokens for largest models
• Scaling Impact: As AI adoption grows, cumulative environmental impact becomes significant

Business & Environmental Benefits

Benefit	Impact
Cost Reduction	More sustainable models are typically 5-50x cheaper
Regulatory Compliance	EU AI Act and sustainability reporting requirements
Corporate Responsibility	ESG goals and stakeholder expectations
Performance	Efficient models often have faster response times

---

How We Calculate Sustainability

CO₂ Emissions Formula

CO₂ (grams) = (Tokens ÷ 1000) × CO₂_per_1k_tokens

Data Sources:

• Artificial Analysis benchmark data
• Model provider specifications
• Industry research on AI carbon footprints

Factors Considered:

• Model size (parameters)
• Hardware efficiency (DGX A100, H100, H200, TPU)
• Data center PUE (Power Usage Effectiveness)
• Carbon Intensity Factor (CIF) by region

Water Consumption Formula

Water (liters) = (Tokens ÷ 1000) × Water_liters_per_1k_tokens

Factors:

• Direct cooling water usage
• Indirect water from power generation
• Data center location and climate

Sustainability Rating System

Models are rated on a percentile-based scale:

Rating	Percentile	CO₂ Range	Description
A+	Top 20%	< 1.0 g/1k	Most sustainable
A	20-40%	1.0-2.0 g/1k	Very sustainable
B	40-60%	2.0-5.0 g/1k	Moderate
C	60-80%	5.0-10.0 g/1k	Higher impact
D	Bottom 20%	> 10.0 g/1k	Highest impact

---

Complete Model Sustainability Reference

All Models Ranked by CO₂ Efficiency

Rank	Model	CO₂/1k (g)	Water/1k (L)	Cost/1k ($)	Rating
1	Amazon Nova Lite	0.10	0.07	0.0005	A+
2	Gemini 2.0 Flash Lite	0.12	0.15	0.0007	A+
3	Gemini 2.5 Flash Lite	0.15	0.18	0.0008	A+
4	Gemini 2.0 Flash	0.45	0.54	0.002	A+
5	Amazon Nova Pro	0.50	0.35	0.003	A+
6	GPT-4.1	0.56	0.44	0.012	A+
7	Gemini 2.5 Flash	0.56	0.66	0.0025	A+
8	GPT-4.1 Mini	0.59	0.46	0.002	A+
9	GPT-4o Mini	0.64	0.46	0.0015	A+
10	Claude 3 Haiku	0.64	0.34	0.0008	A+
11	Claude Haiku 4.5	0.78	0.41	0.001	A+
12	O3 (Beta)	0.99	0.78	0.006	A
13	GPT-4o	1.17	0.88	0.010	A
14	Claude 3.7 Sonnet	1.18	0.62	0.015	A
15	Claude Sonnet 4	1.18	0.62	0.018	A
16	Claude Sonnet 4.5	1.20	0.63	0.018	A
17	Gemini 2.5 Pro	1.54	1.84	0.010	A
18	O4 Mini (Beta)	5.13	4.04	0.002	B
19	GPT-5 Mini	7.75	6.10	0.005	B
20	GPT-5	13.78	17.69	0.020	D
21	GPT-5.1 (Beta)	13.78	17.69	0.025	D

---

Relatable Equivalents

To make environmental impact tangible, we convert metrics to everyday equivalents:

CO₂ Equivalents

CO₂ Amount	Equivalent
1 gram	0.004 km driving
10 grams	Charging a smartphone
100 grams	1 dishwasher cycle
1 kg	4 km driving

Water Equivalents

Water Amount	Equivalent
0.1 liters	1/5 glass of water
0.5 liters	1 water bottle
1 liter	2 water bottles
5 liters	1 minute shower

Example: 10,000 Requests (500 tokens each)

Model	CO₂	Equivalent	Water	Equivalent
GPT-5	68.9 kg	275 km driving	88.4 L	18 min shower
GPT-4o Mini	3.2 kg	13 km driving	2.3 L	5 water bottles
Amazon Nova Lite	0.5 kg	2 km driving	0.35 L	1 glass water

---

Best Practices for Sustainable AI

1. Right-Size Your Model

Rule of thumb: Use the smallest model that meets quality requirements.

Simple tasks → Lite/Haiku models (A+)
Standard tasks → Mini models (A+)
Complex tasks → Pro/Full models (A/B)
Critical tasks → Premium models (C/D) - only when necessary

2. Optimize Token Usage

Reduce input tokens:

• Use concise prompts
• Remove unnecessary context
• Use efficient prompt templates

Reduce output tokens:

• Set appropriate max_tokens limits
• Request concise responses
• Use structured output formats

3. Implement Smart Routing

Route requests based on complexity:

function selectModel(complexity) {
  switch(complexity) {
    case 'simple':
      return 'amazon-nova-lite';     // A+ rating
    case 'standard':
      return 'gpt-4o-mini';          // A+ rating
    case 'complex':
      return 'gpt-4.1';              // A+ rating
    case 'critical':
      return 'gpt-5';                // D rating - use sparingly
  }
}

4. Cache and Batch

• Cache responses for repeated queries
• Batch similar requests to reduce overhead
• Use embeddings for semantic caching

5. Monitor and Report

Track sustainability metrics:

• Total CO₂ emissions per day/week/month
• Water consumption trends
• Cost vs. sustainability correlation
• Model usage distribution

---

Model Selection by Sustainability Priority

🌱 Sustainability-First (A+ Only)

Best for organizations prioritizing environmental impact:

Model	CO₂/1k	Use Case
Amazon Nova Lite	0.10g	Simple tasks, high volume
Gemini 2.0 Flash Lite	0.12g	Fast responses
GPT-4.1	0.56g	Complex tasks with sustainability
GPT-4o Mini	0.64g	General purpose
Claude 3 Haiku	0.64g	High-volume chatbots

⚖️ Balanced Approach (A+ and A)

Good sustainability with broader capability:

Model	CO₂/1k	Use Case
GPT-4o	1.17g	Multimodal tasks
Claude Sonnet 4	1.18g	Complex reasoning
Gemini 2.5 Pro	1.54g	Research, analysis

🎯 Quality-First (Accept Higher Impact)

When quality is paramount:

Model	CO₂/1k	Use Case
GPT-5	13.78g	Most complex tasks
GPT-5.1	13.78g	Cutting-edge features

---

ROI Calculator: Sustainability Switch

Scenario: 100,000 requests/month

Current State: GPT-5 for all requests

• Monthly CO₂: 689 kg
• Monthly Water: 884 L
• Monthly Cost: $1,000

After Optimization: Hybrid approach

• 70% GPT-4o Mini → 224 kg CO₂
• 20% GPT-4.1 → 56 kg CO₂
• 10% GPT-5 → 69 kg CO₂

Results:

Metric	Before	After	Savings
CO₂	689 kg	349 kg	49% reduction
Water	884 L	450 L	49% reduction
Cost	$1,000	$220	78% reduction

---

Data Sources & Methodology

Calculation Methodology

Our sustainability data is based on:

1. Artificial Analysis Benchmarks: Real-world inference measurements
2. Hardware Specifications:
   • NVIDIA DGX A100/H100/H200
   • Google TPU v5e/v6e
3. Data Center Metrics:
   • PUE (Power Usage Effectiveness): 1.09-1.14
   • CIF (Carbon Intensity Factor): 0.231-0.34 gCO2e/Wh

Assumptions

• Inference only (not training)
• Standard 300-token query baseline
• Average data center efficiency
• Regional carbon intensity averages

Limitations

• Actual values vary by:
  • Geographic location
  • Time of day (grid mix)
  • Specific hardware configuration
  • Workload patterns
• Values are estimates; actual impact may differ ±20%

---

Additional Resources

• Salesforce Net Zero Cloud
• AI Carbon Footprint Calculator
• Artificial Analysis Benchmarks
• Green Software Foundation

---

Key Takeaways

1. Small changes = Big impact: Switching from GPT-5 to GPT-4o Mini reduces CO₂ by 95%
2. Cost and sustainability align: Cheaper models are usually more sustainable
3. Right-size your models: Use premium models only when quality demands it
4. Monitor and optimize: Track your AI carbon footprint over time
5. Hybrid approach wins: Route requests to appropriate models based on complexity

---

Building sustainable AI isn't just good for the planet—it's good for your bottom line.

---

Last Updated: January 2025 Data Version: Based on models.json with Artificial Analysis benchmark data