Free AI carbon calculator · No signup needed · Method 2026.11
AI Carbon Footprint Calculator
How much CO₂ do your AI API calls produce?
Estimate the electricity and emissions of ChatGPT, Claude and Gemini usage per call, per successful task and per month. Compare models for the same work and see what drives the result.
Conversations, documents, tickets or agent runs your product completes.
Agents and chat turns usually call the model several times per task.
Prompt, instructions, retrieved context and history. 1,000 tokens ≈ 750 words.
Generated text, including reasoning. The biggest driver of energy.
%
Share of input tokens read from the provider's prompt cache.%
Calls that do not contribute to a successful task.Estimated monthly emissions
19.38 kg CO₂e
0.19 g per successful task · 0.059 g per model call
Low 6.46 kgHigh 38.77 kg
Retries and failed calls add 1.55 kg and $65.54 a month. Cutting them lowers carbon and cost together.
Claude Haiku 5.5 would emit about 73% less for this workload and cost $710 less a month. Check quality before switching.
Where each call's emissions come from
- Generating output58%
- Reading input20%
- Cache reads2%
- Data-centre cooling and power (PUE)7%
- Hardware manufacturing13%
Google Cloud Iowa · grid 413 gCO₂e/kWh · PUE 1.09. Counting the provider's clean-energy purchases (market-based): 5.87 kg.
Same workload on every model
Monthly CO₂e (bar shows the likely range) and list-price cost. Models on another cloud use that cloud’s default region. Select a model to use it above.
Estimates use method 2026.11 and list prices as of 2026-10-30 (standard tier, no batch or long-context pricing). Model sizes for closed models are not published, so each is placed in a size tier. How this is calculated
Calculations run in your browser. Nothing you enter is sent or stored. 326.1K model calls a month in this estimate.
Self-hosted models: kWh × PUE × grid
If you serve models on your own hardware, you can measure the electricity directly. Multiply it by your facility’s power usage effectiveness (PUE) and your grid’s carbon intensity, then divide by the requests served. We add a 15% allowance for manufacturing the hardware, as the SCI specification requires embodied emissions.
Worked examples
These illustrative workloads use the same engine as the calculator, at the provider’s default region. They are not measurements of any customer.
Support chatbot on Gemini 3.8 Flash
A customer-service assistant answering questions with a knowledge base. 100,000 tasks a month, 3 calls per task, 2,500 input and 350 output tokens per call, 40% cached and 8% retries.
| Metric | Result |
|---|---|
| Emissions (likely range) | 19.38 kg (6.46 kg – 38.77 kg) |
| Per successful task | 0.19 g |
| Electricity incl. data-centre overhead | 40.8 kWh |
| Of which retries and failures | 1.55 kg |
| Model cost at list price | $819 |
Coding assistant on Claude Opus 5.5
An agent that reads a codebase, edits files and runs checks. 20,000 tasks a month, 8 calls per task, 12,000 input and 900 output tokens per call, 70% cached and 10% retries.
| Metric | Result |
|---|---|
| Emissions (likely range) | 69.22 kg (23.07 kg – 138.44 kg) |
| Per successful task | 3.46 g |
| Electricity incl. data-centre overhead | 164.8 kWh |
| Of which retries and failures | 6.92 kg |
| Model cost at list price | $6,059 |
Research agent on GPT-6 Astra
A multi-step agent that searches, reads and writes a report. 5,000 tasks a month, 20 calls per task, 6,000 input and 700 output tokens per call, 50% cached and 15% retries.
| Metric | Result |
|---|---|
| Emissions (likely range) | 83.41 kg (27.81 kg – 166.83 kg) |
| Per successful task | 16.68 g |
| Electricity incl. data-centre overhead | 191.3 kWh |
| Of which retries and failures | 12.51 kg |
| Model cost at list price | $8,000 |
Self-hosted: 1,550 kWh a month
A team’s GPUs draw 1,550 kWh of IT electricity a month in a facility with a PUE of 1.25 on a 600 g/kWh grid, serving 100,000 requests. Facility electricity is 1,937.5 kWh and operational emissions are 1.16 t, or 11.63 g per request. With the 15% hardware allowance the total is 1.34 t, or 13.37 g per request.
How the API estimate is calculated
Energy per call = (output tokens + input tokens × 0.08 + cached tokens × 0.01) × the model tier’s energy per token × PUE. CO₂e = energy × grid carbon intensity × (1 + 0.15 for hardware manufacturing).
Energy per token
Closed models do not publish their size, so each is placed in a small, medium, large or frontier tier. Each tier uses the median energy per output token of open models of that size on the ML.ENERGY Leaderboard, then adds host CPU, memory and idle capacity using the shares Google measured for Gemini in production. The low, central and high values reflect efficient, typical and lightly batched serving. Reasoning tokens count as output.
| Tier | Low | Central | High |
|---|---|---|---|
| Frontier | 0.472 | 1.415 | 2.830 |
| Large | 0.211 | 0.634 | 1.268 |
| Medium | 0.080 | 0.239 | 0.478 |
| Small | 0.023 | 0.070 | 0.139 |
Data centre and grid
PUE comes from each hosting cloud’s published figure: Google Cloud 1.09, AWS 1.15 and Azure 1.185. OpenAI is assumed to run on Azure, Anthropic on AWS and Google on Google Cloud. Grid carbon intensity is the published average for the chosen region; the provider default is used otherwise.
What is included
Included: inference for the calls you make, data-centre cooling and power distribution, idle capacity, and a 15% hardware-manufacturing allowance. Excluded: model training, network transfer, your own application servers and end-user devices. The result is an estimate, not a certified footprint. Inside Ganivra, providers that report energy directly replace the estimate.
Sources
- ML.ENERGY Leaderboard v3.0, Text Conversation task (energy-optimal configurations and per-model configuration sweeps) — ML.ENERGY Initiative, University of Michigan, v3.0, updated 2026-09-23
- The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization (NeurIPS 2025 Datasets & Benchmarks) — Chung et al., v2
- Measuring the environmental impact of delivering AI at Google Scale — Google, August 2025
- Google data centers: efficiency (fleet-wide trailing annual PUE) — Google, 2024 and 2025 fleet average 1.09
- 2024 Amazon Sustainability Report, AWS summary (global PUE) — Amazon, 2024
- Cloud Carbon Footprint methodology: PUE coefficients and Appendix II grid emissions factors (EPA eGRID2023, EEA, Google, Carbon Data Intelligence, carbonfootprint.com) — Cloud Carbon Footprint (Thoughtworks), retrieved 2026-10-30
- Software Carbon Intensity (SCI) Specification, ISO/IEC 21031:2024 — Green Software Foundation, ISO/IEC 21031:2024
- Greenhouse gas emissions from a typical passenger vehicle — US EPA (car comparison, about 400 g CO₂ per mile)
AI carbon footprint frequently asked questions
How much CO₂ does one ChatGPT, Claude or Gemini call produce?
A short text call to a small, fast model is usually well under a gram of CO₂e. A long answer from a frontier model can be several grams. Google measured a median Gemini text prompt at 0.24 Wh and 0.03 g CO₂e (market-based) in 2025. Total output tokens, model size and the grid powering the data centre matter most.
Why does the calculator show a range rather than one number?
Providers do not publish energy per request for their closed models, so each model is placed in a size tier and measured with open models of that size on the ML.ENERGY benchmark. The range covers efficient and inefficient serving (batch size and hardware). The central figure is what we report; the range shows how uncertain it is.
Why do output tokens matter more than input tokens?
The model generates output one token at a time, so each output token costs a full pass through the model. Input is processed in parallel and costs roughly a twelfth as much per token. Cached input costs less again. Shorter answers and fewer reasoning tokens are the most effective way to cut emissions.
What is the difference between location-based and market-based?
Location-based uses the average carbon intensity of the local grid. Market-based also counts the provider’s clean-energy purchases. The GHG Protocol asks companies to report both. We publish a market-based figure only where the provider discloses one (Google Cloud); elsewhere it equals the location-based figure.
Why include retries and failed calls?
A retry or a failed run uses the same electricity as a successful one but produces nothing. We report emissions per successful task, so waste raises the figure. Reducing retries, loops and failed agent runs lowers carbon and cost at the same time.
Is this a certified carbon footprint?
No. It is an estimate aligned with the Green Software Foundation’s Software Carbon Intensity specification (ISO/IEC 21031:2024) and suitable for GHG Protocol Scope 3, Category 1 reporting of purchased AI services. It covers model inference, data-centre overhead and a 15% hardware-manufacturing allowance. It excludes model training, your own servers and end-user devices.
I run my own GPUs. Which calculator should I use?
Use the self-hosted calculator. Enter the electricity your servers actually used, your facility PUE and your grid factor. That metered approach is more accurate than any per-token estimate. Use the API calculator for providers that do not share energy data.
Is my data stored?
No. Both calculators run entirely in your browser. Nothing you type is sent to Ganivra. A shared link only contains the numbers in its address.
Measure real usage, not estimates of it
Ganivra meters every AI API call your product makes and attaches its emissions, cost and data quality. See CO₂e per workflow, per customer and per successful task, and export SCI-aligned statements for customer and ESG reporting.
To see what each successful AI outcome costs, use the AI Unit Economics Calculator.