Skip to content
Get in touchFR
Updated in November 2026
Photo credit
Explained

Which AI should you choose for the smallest footprint?

The same query uses 0.05 to 2.3 Wh depending on model size and thinking mode, whatever the brand. The figures published by each provider, simple habits and the AI tools marketed as eco-friendly.

Guillaume Pakula
By Guillaume Pakula, co-founder of Celsius. Since 2020, he has helped 80+ organisations with their Bilan Carbone® and climate strategy.
June 2026
Updated November 2026 · 9 min
The type of model matters far more than the brand: a query costs about 0.05 Wh on a small model, 0.3 Wh on an assistant's standard model and 2.3 Wh on a large model in thinking mode, according to Projet Celsius's AI Impact Database. ChatGPT, Claude, Gemini and Mistral AI all offer these ranges, and only Google and Mistral AI publish a figure with its method. Choosing the smallest model that does the job counts for more than switching assistant.
Key takeaways
  • 1The model matters more than the brand: a query uses 0.05 to 1.2 Wh depending on the model.
  • 2Thinking mode doubles energy use on average, and far more on coding tasks.
  • 3Only Google and Mistral AI publish a figure with its method; Anthropic publishes nothing.
  • 4A 5-second video clip uses as much energy as 330 queries to an AI.

In France, 48% of people aged 12 and over used generative AI in 2025, up from 20% in 2023, according to the Digital Barometer produced by the Crédoc research centre for Arcep (the French electronic communications regulator). Each query weighs little, but their number shows up in data centres, the buildings where the servers run: their global electricity consumption is projected to rise from about 415 TWh in 2024 to nearly 950 TWh in 2030 in the base case of the International Energy Agency (IEA).

Choosing a model

Electricity used by the same task, from small to large model

Choose a use: each row shows what the servers consume to do it once.

For which use?

Small model

HaikuFlash-LiteSmall

0.05 Wh

0.023 gCO2e

the lightest in the range, or 380 times for 1 phone charge

Standard model

InstantSonnetFlash

0.3 Wh

0.14 gCO2e

the default model, or 63 times for 1 phone charge

Large model

OpusPro

1.2 Wh

0.55 gCO2e

the most powerful, or 16 times for 1 phone charge

For a query and a half-page answer, the large model uses about 24 times more than the small one. Thinking mode multiplies energy use by 1.9 on average.

Projet Celsius AI Impact Database, November 2026: electricity for the servers and the data centre, global kilowatt-hour at 458 gCO2e (Ember). Providers do not publish the size of their models: the names place each model within its range. Thinking mode: PEReN, 22 open models. Phone charge: 19 Wh.

The answer to the question in the title therefore lies less in a logo than in a setting. ChatGPT, Claude, Gemini and Mistral AI's assistant each offer a light model, a default model and a more powerful model, and most add a thinking mode. The figures in this article come from the AI Impact Database (in French), Projet Celsius's database of AI impacts, in watt-hours (Wh): a phone charge takes 19.

1Which AI uses the least energy, model by model?

For a query and an answer half a page long, a standard-sized assistant uses about 0.3 Wh, or 0.14 gCO2e (grams of CO2 equivalent) with the global average electricity mix. That is the value on which a Google measurement, 2 estimates and a statement from OpenAI converge, as set out in our article on the energy used by a ChatGPT query, roughly the same as a Google search. The same query costs about 0.05 Wh on a small model and 1.2 Wh on a large model.

Small, standard or large: model size

A larger model performs more computation for every word it writes, and electricity follows computation. The 4 main assistants rank their models in the same way, from the lightest to the most powerful:

  • At Anthropic, Claude comes as Haiku, designed for fast, high-volume tasks, then Sonnet, Opus and Fable, reserved for the most demanding reasoning.
  • At Google, Gemini offers Flash-Lite, the fastest and cheapest, then Flash and Pro, which the app presents as the model that thinks through complex topics.
  • At OpenAI, ChatGPT answers by default with a fast model, Instant, and can switch to its Thinking mode, which reasons before answering.
  • At Mistral AI, the Le Chat assistant, renamed Vibe in 2026, relies on a range that runs from Mistral Small, the lightest, to Mistral Large.
Screenshot of the Gemini app on a phone: the "Choose your model" menu offers "Fast, Answers quickly" and "Thinking with 3 Pro, Thinks through complex topics"
The model menu in Gemini, in December 2025: "Fast" answers directly, "Thinking" reasons before answering. For a simple query, choosing the first avoids doubling electricity use on average; ChatGPT, Claude and Vibe have the same menu. Screenshot by Orwell reader, Wikimedia Commons, CC BY-SA 4.0, cropped

None of these providers publishes the size of its models. Classing them as small, standard and large remains an assumption based on their place in the range and on their price, which partly tracks the amount of computation: according to its public price list, Anthropic charges developers 40 times less for Haiku than for Opus. The values in the opening chart are those of the AI Impact Database for a typical size. At Anthropic, the choice between the 2 large models is covered in Fable 5.2 or Opus 5.5.

An academic study gives an order of magnitude model by model. In 2025 its authors estimated the electricity used by a short query followed by an answer of about 200 words, based on each model's public response speed and the likely hardware in the data centres: these are estimates, not measurements by the providers. According to "How Hungry is AI?" (Jegham et al., November 2025 version):

  • GPT-4.1 nano, OpenAI's smallest model in the study, uses 0.21 Wh per query.
  • GPT-4o, ChatGPT's default model until August 2025, uses 0.42 Wh, or 0.19 gCO2e with the global average electricity mix.
  • Anthropic's Claude 3.7 Sonnet uses 0.95 Wh, almost as much as Claude 3.5 Haiku (0.98 Wh), even though Haiku was the fast model in the range at the time.
  • o3, OpenAI's reasoning model, uses 1.2 Wh, and o4-mini with high reasoning effort 3.6 Wh.

Almost all of these 2025 models have since been replaced. The Haiku case shows that a "fast" model name does not guarantee an economical query, and that the method sees neither the actual size of the models nor the exact machines. The orders of magnitude match those of the AI Impact Database: a few tenths of a watt-hour for a standard model, more than 1 Wh for a large model or a model that thinks.

Thinking mode: energy use doubled on average

Thinking mode has the model write a draft you do not see before its answer. Across 22 open models measured by PEReN, the French digital regulation expertise centre, it increases energy use by 92% on average, and by 849% on average on programming exercises, according to Arcep's May 2026 report. A large model in thinking mode therefore uses about 2.3 Wh for a query, 46 times more than a small model answering directly.

Within the same assistant, the choice of model and of thinking mode changes the footprint of a query more than the choice of assistant.

2What do OpenAI, Anthropic, Google and Mistral publish?

Only 2 providers have published a figure together with its method, and no figure compares directly with the others: they count neither the same thing nor the same query.

A measurement, a sentence, an assessment and a silence

Google measured the median query to Gemini (in French) in production: 0.24 Wh and 0.26 mL of water, the equivalent of 5 drops, according to its August 2025 paper. The same paper states that the energy of this query was cut 33-fold in one year. It is the only production measurement published by a provider, and no third party has audited it.

OpenAI sticks to a single sentence from its chief executive, Sam Altman: an average query would use about 0.34 Wh, according to his June 2025 blog post, with no method or scope. Anthropic publishes no figure for Claude. In July 2025, Mistral AI published the first life cycle assessment of a large model, carried out with ADEME (the French Agency for Ecological Transition): 1.14 gCO2e and 45 mL of water for one answer from its assistant, server manufacturing included.

Published figures

What Google, OpenAI, Anthropic and Mistral AI publish

The figure given for a query to each assistant, and what it counts.

GoogleGemini
A production measurementAugust 2025

0.24 Wh and 0.26 mL of water per median query

Chips, servers and data centre; method published

OpenAIChatGPT
A sentence from its chief executiveJune 2025

0.34 Wh and 0.32 mL of water for "an average query"

No method or scope published

AnthropicClaude
NothingNovember 2026

No figure neither per query nor per model

No model size published either

Mistral AILe Chat (Vibe)
A life cycle assessmentJuly 2025

1.14 gCO2e and 45 mL of water per half-page answer

Full life cycle, server manufacturing included

These figures cannot be ranked against each other: electricity or carbon, short query or long answer.

Sources: Google, arXiv 2508.15734; Sam Altman, June 2025; Mistral AI, July 2025; Anthropic; Projet Celsius AI Impact Database.

Which AI uses the least water?

Google's 0.26 mL counts only the water evaporated to cool its data centres. Mistral AI's 45 mL counts the water used across the whole life cycle, including chip manufacturing and electricity generation. The gap comes from the scope, and does not say which of the 2 assistants uses less water. Our article on the water AI uses details these 3 kinds of water, and the one on the stages at which AI harms the environment the weight of manufacturing.

3Which habits reduce your AI footprint?

Model size is not the only lever. The nature of the task makes electricity use vary far more: from 0.13 Wh for the summary generated at the top of a search to 670 Wh for one hour of a coding agent (in French), a program that chains dozens of steps on its own.

AI uses

Electricity used by common AI tasks, from text to video

One row per use, on a scale where each mark is worth 10 times the previous one.

Ask a question0.3 Wh
Create an image3.2 Wh
Query a 30-page PDF4.7 Wh
Transcribe and summarise a 1-hour meeting30 Wh
Run a deep research task99 Wh
Create a 5-second video clip99 Wh
Have an agent code for 1 hour670 Wh

A 5-second video clip uses as much as 330 queries, one hour of a coding agent as much as 2,200.

Projet Celsius AI Impact Database, November 2026, standard model without thinking mode. Phone charge: EPA.

5 habits reduce the footprint without giving up the tool, from the simplest to the most effective depending on use:

  • Start with your assistant's lightest model, and move up only if the answer is not good enough: for a query, the gap runs from 0.05 to 1.2 Wh.
  • Leave thinking mode off for simple questions, a fact search or an email: it doubles energy use on average, for a gain in quality that varies from task to task according to Arcep.
  • Attach a document only when the answer depends on it: querying a 30-page PDF costs about 4.7 Wh, as much as 16 queries, because the model reads the whole text.
  • Keep images and video for real needs: a generated image takes about 3.2 Wh, and a 5-second clip nearly 100 Wh, as much as 330 queries.
  • Open a new conversation when you change subject: with each message, the model rereads the earlier exchange, and 10 exchanges already use about 3.5 Wh.

Our article on 10 habits ranked by effectiveness quantifies them over a year of use, and those on the footprint of an AI image and of an AI-generated video cover the heaviest uses.

4Ecosia, Mistral: is there such a thing as green AI?

No AI is green in itself. Some combine more favourable conditions: a small model, low-carbon electricity, published figures. Our comparison of AI tools marketed as eco-friendly covers Ecosia AI, Infomaniak's Euria, GreenPT and Vibe.

Ecosia AI, a small Mistral AI model

The Ecosia search engine runs its AI on Mistral AI models, including Mistral Small 3.2 for the summaries at the top of results, according to its help centre. In a blog post of 2 July 2026, it claims to produce more clean energy than its AI uses, without publishing a figure. Its AI features can be switched off in the settings, and it estimates the impact of each request with the EcoLogits method, which leaves out training, networks and the user's device.

Mistral AI and the country where the servers run

Mistral AI remains the only provider to have published a life cycle assessment, reviewed by 2 engineering consultancies. It does not say that its assistant emits less than ChatGPT or Gemini: its 1.14 gCO2e include server manufacturing, which the other figures leave out.

A former brick port warehouse bearing the words Magasin Vauban, flanked by technical units, in Strasbourg
The Magasin Vauban, a former warehouse in the port of Strasbourg converted into a data centre, in 2013. The country where the servers run sets the carbon of each kilowatt-hour: an example of a French site, unrelated to the assistants mentioned. Photo by Michael Kauffmann, Wikimedia Commons, CC BY 3.0

The servers' location matters too. Producing 1 kWh emits 41 gCO2e in France, against 384 in the United States and 458 on the global average, according to the Ember data used by the AI Impact Database: the same query therefore emits about 11 times less in France. Yet no major assistant states where the answer to a given query is computed.

5Does AI weigh heavily in a company's footprint?

In a company, the same reasoning applies at a larger scale, and it is intensive uses that weigh. One hour of a coding agent uses about 670 Wh, or 310 gCO2e with global electricity, the equivalent of 35 phone charges. Setting a light model as the default in internal tools, then reserving large models and thinking mode for the teams that need them, is the first lever, and it costs nothing.

Our article on AI in a company's Bilan Carbone®, the French carbon accounting method, explains how to estimate this item and what it represents. The AI carbon footprint calculator gives an order of magnitude, and the method for including AI in your Bilan Carbone® is covered separately, as is a company's digital footprint.

For us at Projet Celsius, the useful question in front of an assistant is "what is the smallest model that does this job properly?". The providers do not publish enough to measure each brand on the same basis: our environmental ranking of AI assistants (in French) compares them on 4 criteria, from figures of different kinds. Each of them does, however, let you choose the model size and thinking mode, and these 2 settings make the electricity used by a query vary from 1 to nearly 50. We recommend that individuals start with the lightest model, and that companies make it the default setting in their tools.

6Key takeaways

  • The type of model matters more than the brand: a query costs about 0.05 Wh on a small model, 0.3 Wh on a standard model and 1.2 Wh on a large one.
  • Thinking mode increases energy use by 92% on average according to PEReN's measurements, and by 849% on average on programming exercises.
  • Google publishes a measurement (0.24 Wh), OpenAI a sentence (0.34 Wh), Mistral AI a life cycle assessment (1.14 gCO2e), and Anthropic nothing for Claude.
  • The task weighs more than the assistant: a 5-second video clip uses as much as 330 queries, one hour of a coding agent as much as 2,200.
  • No AI is green in itself: a small model, without thinking mode, on servers powered by low-carbon electricity remains the most economical choice.

To place these figures among AI's wider impacts, see all the environmental impacts of AI, AI's electricity consumption worldwide and should you stop using AI?

Further resources

Frequently asked questions

or: guillaume@projetcelsius.com

More articles to read

View all →
French environmental cost label for clothing: what changed on 1 October 2026
Regulation

French environmental cost label for clothing: what changed on 1 October 2026

6 min read
What is the French textile eco-score?
Who displays it, and has it become mandatory?
Regulation

What is the French textile eco-score? Who displays it, and has it become mandatory?

12 min read
Removable batteries and battery passports: who is affected in 2027?
Regulation

Removable batteries and battery passports: who is affected in 2027?

14 min read

A senior consultant reads every message

Tell us where you stand and what is pressing, and we will reply within 24 working hours with a first view on how to proceed.

Write to a consultant →
Our toolkit

A full range of free assessment tools

Cost estimator, obligation and eligibility checkers, footprint calculator.

Estimate costs and impacts
Ref. 2026·BC·0184REPORT · FINANCIAL YEAR 2026Bilan Carbone®complete 1·2·3Presentation to the Executive Committee · 2027 action planCONTENTS01Executive summaryp. 402Scope and methodologyp. 1203Emissions by scopep. 24042027 action planp. 5605Methodological appendicesp. 78VOLUME84 pages · ConfidentialSTANDARDISO 14064 · GHG ProtocolQUOTATION · BC.2026.0184Issued 05·05·2026 · Valid for 30 daysCelsius consultancy · Paris 3rd arrondissementINDICATIVE COST ESTIMATEYour companyMid-sized company · 180 employees · CSRD 2027SERVICEBilan Carbone®Scope 1·2·3 · 13 weeksBPIFRANCE FUNDING · - 60%INDICATIVE RANGE EXCL. VAT€16,800to €22,400PROJECT DETAILS01 · SCOPING3 weeks~€4,20002 · MODELLING8 weeks~€9,80003 · PRESENTATION2 weeks~€4,800Indicative estimate · Refined after scopingSIRET 891 234 567 00012

Bilan Carbone® cost estimator

Your price range in 1 minute, based on 2026 market rates and public funding.

1 minStart
Estimate costs and impacts
AI FOOTPRINT · MONTHLY USEYOUR TOTAL FOOTPRINT4.2 kgCO₂eq for 1,000 typical prompts · May 2026MODEL COMPARISON · 1,000 TYPICAL PROMPTSGPT-4oOpenAI6.8 kgClaude OpusAnthropic2.1 kgGemini ProGoogle4.5 kgMistral LargeMistral AI · FR1.6 kg

AI carbon footprint calculator

The climate impact of your AI queries, by model and task. Ecologits methodology.

3 minStart
Check an obligation
DIAGDÉCARBON'ACTIONACCELERATE THETRANSFORMATIONOF YOUR COMPANYIN ENERGY ANDTHE ENVIRONMENTOFFICIAL SCHEME · BPIFRANCE × ADEMECHECKER · ELIGIBILITYELIGIBLEYou meet the 3 Diag Décarbon'Action criteria.YOUR FUNDING€10,000 excl. VAT€6,000excl. VATAmount payable by you after the Bpifrance subsidyFUNDING BREAKDOWNBPIFRANCE 40%YOU 60%€4,000€6,000SCHEME SUBSIDISED BY

Diag Décarbon'Action eligibility

Check in 30 seconds whether your business is eligible for Bpifrance funding covering 40% of a Bilan Carbone® assessment.

30 secStart
Check an obligation
FRENCH REPUBLICMinistry for the Ecological TransitionLEGAL REQUIREMENTGreenhouse GasEmissions AssessmentArticle L. 229-25 of the French Environmental CodeDecree 2022-982 · published 1 July 2022FREQUENCY · 4 YEARSFILING · ADEMECHECKER · BEGES REQUIREMENTREQUIRED720 employees · BEGES requirementPOTENTIAL PENALTY€50,000maximum fine for failing to complete itart. L. 229-25OFFICIAL REFERENCE

BEGES checker

Does the BEGES requirement apply to you? An immediate answer, with the deadline and penalty.

30 secStart
Check an obligation
9:42SCANSCANNING DPP...DPP IDENTIFIEDOrganic cotton T-shirtSKU TX-CB-220 · Batch L-26-04781ESPR COMPLIANT42 DATA POINTS · 5 SECTIONSORIGINIndia · GOTSOrganic cottonFOOTPRINT5.2 kg CO₂eqLCA ISO 14040RECYCLABILITY85%Cotton recycling schemeREPAIRABILITY7.5 / 10Accessible partsEND OF LIFETextile recycling scheme (TLC) · Bordeaux/FRRefashion-approved recovery operatorEU · ESPR 2024/178112.05.2026

DPP checker

Is your product covered by the Digital Product Passport?

1 minStart
Check an obligation
COSTENVIRONMENTAL386POINTS257PER 100 GOfficial methodEcobalyse, v7.0.0CHECKER · ENVIRONMENTAL LABELLINGELIGIBLETextiles · voluntary labellingENVIRONMENTAL COST386 ptsT-shirt 150 g · 257 pts per 100 gExample: Ministry for the Ecological Transition FAQ

Textile environmental labelling 2026

Voluntary labelling, but third parties can publish it from October 2026: where do you stand?

1 minStart