About the toolbox
Every prompt has a price. This toolbox gathers 25 tools for tracking the energy use and carbon footprint of generative AI, from code libraries and leaderboards to web calculators and research methods, so developers, researchers and everyday users can pick the right instrument for their needs.
Developed by Janet Rafner and Carina Gibat (2026) and maintained by the Center for Hybrid Intelligence, Aarhus University. Funded by Aarhus University iClimate, the Carlsberg Foundation and Erasmus+. For more information, contact: janetrafner@mgmt.au.dk
How to cite: Janet Rafner and Carina Gibat. 2026. Every Prompt Has a Price: A Toolbox for Tracking the Environmental Cost of GenAI Use. In Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26). Association for Computing Machinery, New York, NY, USA, Article 264, 1–5. https://doi.org/10.1145/3772363.3799058
Code libraries and trackers (10)
Install these in your own code or servers to measure energy and emissions while you train or run models.
Zeus Project (2025)
By Jae-Won Chung / Zeus Team · For Developers · Measures Energy consumption · Phase Training and inference
Polls system resources to measure CPU/GPU time and energy, as a command line tool or inside Python code, and exports metrics to Prometheus. Strength: very versatile. Weakness: energy only, no emissions.
CodeCarbon
By CodeCarbon contributors · For Developers, researchers · Measures Power consumption, CO2 estimate · Phase Training and inference
Measures CPU/GPU energy of Python code or running processes and converts it to CO2 using country-specific carbon intensity, or real-time data with an Electricity Maps API token. Strength: easy-to-understand emission equivalents and infrastructure comparisons.
Green Algorithms 4 HPC
By Green Algorithms · For Developers · Measures Power consumption, CO2 estimate · Phase Training and inference
Reads workload manager logs on SLURM-based HPC servers (adaptable to others) and reports carbon footprint, energy, compute, memory efficiency and the impact of failed jobs. Strength: reports both estimates and measurements, which makes it transparent.
Carbontracker (2020)
By Anthony et al. · For Developers, researchers · Measures Power consumption, CO2 estimates and predictions · Phase Training and inference
Uses hardware type, hours, cloud provider and region to predict emissions, and can stop training once a CO2 threshold is crossed. Strength: forecasts the full run from early training and lets you actively cut impact.
Eco2AI
By Budennyy et al. · For Developers · Measures Power consumption, CO2 estimate · Phase Training and inference
Monitors CPU and GPU energy for any Python script and estimates emissions with regional emission coefficients. Results are logged to a local file.
CUMULATOR (2020)
By Trebaol, Hartley, Jaggi and Shokri Ghadikolaei · For Researchers, developers · Measures Power consumption, CO2 estimate · Phase Training
Tracks GPU load and data traffic using fixed assumptions (1 GPU hour = 112 gCO2eq, 1 GB of data center traffic = 31 gCO2eq). Weakness: older assumptions and lower-bound estimates only.
eco4cast (2023)
By Tiutiulnikov et al. · For Developers, researchers · Measures Carbon footprint of neural network training · Phase Training
Schedules training for the hours when grid emissions are lowest, on Google Cloud or locally. Strength: step-by-step guides. Weakness: supports limited regions in the Americas and Europe.
Power Profiler (2023)
By Karim Boubouh, Robert Basmadjian · For Android developers, researchers · Measures Energy use of ML on Android devices · Phase Inference
Open-source platform that monitors voltage, current and CPU usage to track the energy of ML algorithms on Android in real time. Weakness: very niche and not recently updated.
LLMCarbon (2024)
By UnchartedRLab · For Developers · Measures Power consumption, CO2 estimate · Phase Training
Predicts carbon emissions before training, based on the type of LLM, the hardware and the data center’s power usage effectiveness.
PowerAPI (2025)
By Spirals research group, University of Lille and Inria · For Developers, IT administrators · Measures Power consumption · Phase Training and inference
Toolkit for building software-defined power meters that estimate power use in real time from physical meters, processor interfaces and OS counters. Strength: works per process, thread, container or VM. Weakness: complex, not for beginners.
Leaderboards (2)
Compare the energy efficiency of public AI models before you choose one.
AI Energy Score Leaderboard (2025)
By Luccioni et al. (Hugging Face) · For Researchers, general audiences · Measures Relative energy efficiency of AI models · Phase Inference
Benchmarks models on 1,000-task sets on NVIDIA H100 GPUs, tracked with CodeCarbon, across text generation, reasoning, image generation and classification. Note: the score covers energy only, not other environmental impacts.
ML.ENERGY Leaderboard (2025)
By ML.ENERGY Initiative · For Researchers, general audiences · Measures Relative energy efficiency of AI models · Phase Inference
Benchmarks LLMs, multimodal and diffusion models on H100 and B200 GPUs, measures GPU energy with Zeus and maps time-energy trade-offs to suggest optimal setups.
Web calculators and online tools (8)
No installation needed. Enter a few details about your task, model or hardware and get an estimate.
Green Algorithms calculator
By Green Algorithms · For Researchers, general audiences · Measures Carbon emissions · Phase Training and inference
Multiplies estimated energy use (runtime, cores, memory, data center PUE) by regional carbon intensity, adjusting for repeated runs. Strength: includes guidelines and a discussion forum.
HCI GenAI CO2ST Calculator
By Nanna Inie, Jeanette Falk, Raghavendra Selvan · For Researchers, general audiences · Measures Carbon emissions · Phase Inference
Estimates the footprint of AI use by research phase, use type, model and document count, based on local inference tests and Carbontracker data. Strength: tailored to HCI research, with tips to cut impact. Weakness: estimates only, no stated update commitment.
ML CO2 Impact Calculator
By Mila and Element AI (now ServiceNow) · For Researchers, general audiences · Measures Power consumption, CO2 estimate · Phase Inference
Estimates raw and offset emissions from hardware, runtime and cloud provider using region-specific carbon intensity. Weakness: data center PUE is not included automatically, and data may be outdated.
EcoLogits Calculator
By GenAI Impact (non-profit) · For Researchers, general audiences · Measures Electricity, GHG emissions, abiotic resources, primary energy, water · Phase Inference
Estimates the impact of a prompt from the model, hardware and assumed data center location. Strength: broad, holistic view with clear visuals for non-experts. Weakness: data is partly estimated or generalized.
EnergyVis
By Shaikh et al. · For Researchers · Measures Carbon footprint, energy consumption · Phase Training
Interactive tracker that calculates metrics from model, hardware and region; models can be customized or uploaded. Weakness: map covers the USA only and may be too technical for beginners.
Deep Neural Network Energy Estimation Tool
By Tien-Ju Yang, Yu-Hsin Chen, Vivienne Sze (MIT) · For Developers of convolutional neural networks · Measures Energy efficiency · Phase Inference
Online tool for evaluating and designing energy-efficient deep neural networks for embedded processing, with results in about 10 seconds. Strength: detailed, visual output.
Low Carbon GenAI Toolkit
By DGC Green, Directors Guild of Canada, Decarbonade · For General audiences · Measures CO2 estimate · Phase Inference
Estimates emissions for text summaries, images, audio clips or video from output length and number of prompts. Strength: easy to use, with real-life equivalents. Weakness: only four activities and rough estimates.
GAISSALabel
By Duran, Castaño, Gómez, Martínez-Fernández · For Developers, researchers · Measures Energy efficiency · Phase Training and inference
Turns entered or uploaded metrics (for example from CodeCarbon) into an A to E energy label, with improvement suggestions and a plug-in for automatic data collection. Weakness: labels are customizable, so not universal.
Methods, frameworks and algorithms (5)
Research methods for experts who want to measure, model or reduce impact at a deeper level.
NetAdapt
By Yang, Howard, Chen, Zhang, Go, Sze, Adam · For Developers of deep neural networks · Measures Latency, energy consumption · Phase Inference
Algorithm that progressively simplifies a pre-trained network for mobile platforms until a resource budget is met, while maximizing accuracy.
A method to estimate the energy consumption of deep neural networks (2017)
By Yang, Chen, Emer, Sze · For Developers and scientists working with DNNs · Measures Energy efficiency · Phase Inference
Analyses how hyperparameters such as layers and filter sizes affect energy use, focusing on convolutional and fully connected layers, to design efficient networks and mobile apps.
Measuring the environmental impact of delivering AI at Google scale (2025)
By Elsworth et al. / Google · For Experts seeking a comprehensive methodology · Measures Energy, emissions, water · Phase Inference
Four-part methodology covering measurement boundary, energy, emissions and water, and an aggregate metric across the full serving stack. Weakness: focused on Gemini and published by Google, so possible bias.
G-TRACE
By Kausar, Latif, Shahzad, Fatima · For Researchers · Measures CO2 estimate · Phase Inference
Framework that estimates emissions from widely adopted GenAI trends in three steps: trend data collection, workload simulation and CO2 estimation that accounts for regional differences and uncertainty.
SPROUT (2024)
By Li, Jiang, Gadepally, Tiwari · For Researchers · Measures Carbon efficiency of LLMs · Phase Inference
Framework that cuts emissions per inference by working out how short an answer can be without losing quality.
Full data
The complete dataset, including extra notes on each tool, is also available as a spreadsheet.