Calculator · 30 seconds
What is the carbon impact of an AI request ?
Move the slider, add models, compare. EcoLogits method, indicative figures ±40%.
Step 1 · Request length
400tokens
≈Write an email
Indicative figures: catalogue recalibrated on 10 June 2026 against EcoLogits (November 2025 methodology, ±40% intervals). Models marked (i) rely on assumed architectures. These equivalences are educational rather than rigorous measurements.
Step 2 · Models to compare
GPT 5.5Claude Opus 4.8Gemini 3.1 ProMistral Large 3Claude Haiku 4.5
Median footprint per request
0,91g CO2e· between 0,02 g CO2e and 2,6 g CO2e depending on the model
Scope: inference only (request energy + amortised hardware share), excluding training. For a full life cycle assessment, allow around 8 to 10 times more (Mistral LCA: 1.14 g for 400 tokens).
Details by model · 400 tokens · ≈ write an email
Gemini 3.1 Pro2,6 g CO2e
GPT 5.5i1,3 g CO2e
Claude Opus 4.80,91 g CO2e
Mistral Large 3i0,11 g CO2e
Claude Haiku 4.5+ efficient0,02 g CO2e
ReadingWhat to remember
À 400 tokens, Gemini 3.1 Pro emits ×117 more CO₂ than Claude Haiku 4.5. Move the slider: the order can change with prompt length.
To explore further
Shall we calculate your company's full digital footprint?
Beyond personal AI use, we audit your digital scope 3 across the entire boundary: data centres, SaaS, generative AI in production. Open methodology, actionable deliverables.
Our toolkit
A full range of free assessment tools
Cost estimator, obligation and eligibility checkers, footprint calculator.
Frequently asked questions
Frequently asked questions
Model, provider, data centre, training versus inference, inclusion in Bilan Carbone®: the questions we are most often asked about generative AI's carbon footprint.
For the same use, compact models (Claude Haiku, GPT-4o mini, Gemini Flash) consume 5 to 10 times less than XL models (Claude Opus, GPT-5, Gemini Pro). The choice of model has more impact than the choice of provider. Our calculator compares models using the Ecologits methodology (Boavizta). See our AI carbon footprint guide.
Three practical options: use a compact model for simple tasks, avoid redundant requests (save useful answers), choose a data centre in France or Northern Europe (an electricity mix 5 to 10 times less carbon-intensive than in the United States or China, according to ADEME Base Empreinte factors).
The calculator covers inference (your request being processed by the model) and an amortised share of server and GPU manufacture, using the Ecologits methodology. It does not include the model's initial training, which amounts to thousands of tonnes of CO2e for a large model: for a full life cycle assessment, allow around 8 to 10 times more. See the “Why does it seem so low?” modal for more details, or our full AI footprint analysis.
Both figures are correct: they do not measure the same thing. Our calculator estimates inference alone (the energy used by your request, plus an amortised share of hardware manufacture), using the Ecologits method: around 0.1 g for a 400-token response. The LCA published by Mistral in 2025 with Carbone 4 and ADEME covers the full life cycle: amortised model training, server manufacture, data centre ancillary facilities. Result: 1.14 g for the same response, a factor of 8 to 10. It is the same logic as comparing fuel consumed on a journey with the full cost of owning a car, including purchase and maintenance. For an indicative full life cycle figure, therefore multiply our figures by 8 to 10.
The IEA estimates that data centres, driven by AI, will rise from around 1.5% of global electricity in 2024 to nearly 3% in 2030, from 415 to around 945 TWh (Energy and AI). The net effect will depend on uses: if AI accelerates decarbonisation (energy optimisation, climate research), the balance can be positive. If it multiplies entertainment uses, negative. See ADEME sustainable digital technology for the national framework.
The AI footprint is included in the digital scope 3 of your Bilan Carbone®, in the “digital services” category. ADEME emission factors do not yet cover recent AI models: we use proxies or provider data (Google and Mistral have published figures; OpenAI and Anthropic have still published no official data to date). Observed indicative figure: a few kilograms of CO2e per year for an intensive user, less than 1% of a service-sector SME's assessment. See our mandatory scope 3 guide.
Probably a little more, but Google does not publish a separate figure for AI Overview. Per search, the indicative figures remain low: 0.2 to 0.3 Wh for a conventional search, 0.24 Wh for a median Gemini request according to Google's published measurement. A conversation with 10 exchanges on a common model uses around 3 Wh, or 0.2 to 1.2 g CO2e depending on the data centre country. Volume matters: around 14 billion Google searches per day, an increasing share of which generates an AI summary. See our comparison of ChatGPT and Google search.
