Which eco-friendly AI chatbot should you choose in 2026?
What they promise, what you can check

No AI is eco-friendly in itself. To choose an eco-friendly AI, or at least a more frugal one, we use 3 criteria: model size (the software that produces the answers), which changes the energy needed to generate a text about 100-fold; the country where the servers run, because generating electricity emits 9 times less greenhouse gas in France than in the United States, according to our calculation; and what the provider publishes, the part that can be checked.

On 3 October 2026 we read what 4 services publish: Ecosia AI, Euria, GreenPT and Vibe, the name given to Mistral AI's Le Chat in May 2026. Each puts forward an argument of its own: renewable electricity, recovered heat, offsetting or a life cycle assessment, a method that counts impacts from manufacturing to end of life. For none of the 4 did we find all of the following: the model that answers, a public figure per answer, the method behind it and the country of the servers.

1What "eco-friendly" means for an AI, and the criteria that count

Frugality, renewable electricity, offsetting: 3 different promises

The word "eco-friendly" covers 3 promises. Frugality means using less: a smaller model, no video generation, better-cooled buildings. Renewable electricity means buying or generating it. Offsetting means funding projects elsewhere, often forestry ones, that are meant to balance out one's own emissions. The last 2 do not change how much energy an answer consumes.

Electricity described as 100% renewable does not show where a server's power comes from. ADEME (the French Agency for Ecological Transition) writes, about the offers sold to household customers, that on the grid "all green and non-green production is mixed" (our translation). According to the agency, a green offer only guarantees that the supplier buys certificates issued by renewable electricity producers, known as guarantees of origin.

Directive (EU) 2024/825 lists "eco-friendly" and "green" among the generic environmental claims, which are banned when the trader does not specify the claim and cannot demonstrate "recognised excellent environmental performance". It also bans claims of neutrality based on offsetting. Member States had to apply these rules from 27 September 2026.

In France, the bill that transposes these rules was passed by the Senate (the upper house of Parliament) in February 2026 but, according to its legislative file, had not yet been examined by the National Assembly (the lower house) in early October 2026. We express no view on the compliance of any service. Our articles explain banned environmental claims (in French) and how to substantiate an environmental claim (in French).

The 3 criteria we use

Model size. Models cut text into word fragments called tokens. For each token produced, a model with 70 billion parameters or more (its internal settings) uses about 100 times more energy than a small one, according to a 2024 study, revised in 2026 of 9 open models, that is, models that anyone can download freely.

Reasoning mode, in which the model works through intermediate steps, widens the gap further. In the AI Energy Score ranking, reasoning models use 30 times more energy on average than the others, and PEReN, the French state's digital regulation expertise centre, measures up to 92% more on average when this mode is switched on.

The country of the servers. Generating 1 kWh (kilowatt-hour) of electricity emits 41 gCO2e (grams of CO2 equivalent) in France, 39 in Switzerland and 57 in Finland, but 384 in the United States and 589 in Poland, according to 2025 data from the think tank Ember, as reproduced by Our World in Data. These figures cover electricity generated in each country, without taking cross-border exchanges into account. ADEME, which counts electricity consumed, uses 51.9 gCO2e per kWh for France, the value used in the rest of this series. The efficiency of the building matters much less: the energy gap between a very efficient data centre (the building where the servers run) and the sector average is 1.4 times, according to our calculation.

What the provider publishes. Without the model used, a figure per answer, the method behind it and the country of the servers, a promise cannot be checked. This is the approach we take to companies' claims, as set out in our method for avoiding greenwashing without knowing it (in French).

2Ecosia AI, Euria and GreenPT: what each one claims

Ecosia AI: producing more clean energy than its AI uses

The search engine Ecosia launched Ecosia AI in December 2025; its chat is in beta, according to its help centre. In a blog post of 2 July 2026, it says it produces more clean energy than its AI uses. On the pages we read, we found no figure for either the energy it produces or the energy its AI uses.

Ecosia publishes its installed capacities, including a 13 megawatt-peak solar farm in Germany (the maximum output of its panels), held in a joint venture, and €18 million invested in renewable energy projects. It also writes that in 2024 it produced 2 times the electricity used by its searches, a figure that predates the launch of Ecosia AI.

On frugality, Ecosia names its models: Mistral Small 4 for the chat and Mistral Small 3.2 for the summaries at the top of search results, although an older, undated help page still cites an OpenAI model. The service does not generate video, but it does generate images, and the AI can be turned off in 1 click. According to its help centre, Ecosia estimates the impact of each request with EcoLogits, an estimation method also used by the public comparison tool compar:IA, without counting model training, networks or the user's device; we found no quantified results from it. We found no mention of the country of the servers and no price displayed: the launch blog post presents Ecosia as free.

Euria: server heat fed into Geneva's district heating

Euria was launched in December 2025 by the Swiss hosting provider Infomaniak, which presents it as a sovereign and free assistant. According to Infomaniak, it runs on the company's own servers in Switzerland, with open models; Infomaniak cites their families (Mistral, Llama, DeepSeek and Apertus) but not their versions. It also announces certified renewable electricity: in 2024, 38% under the Naturemade Star label, 56% certified hydropower and 6% self-generated solar, according to its impact report.

Infomaniak's environmental argument is that the heat from the servers is fed into Geneva's district heating, which at full capacity is enough to heat up to 6,000 homes in winter and to avoid 3,600 tonnes of CO2 a year, according to its press release.

The 6,000 homes and the 3,600 tonnes describe Infomaniak's Geneva data centre at full load, a level targeted by 2028, while a commercial page from Infomaniak and an article on its blog state the 6,000 homes figure in the present tense. In January 2025, the centre was running at 25% of its capacity, according to the inauguration press release; we found no more recent figure.

Infomaniak writes that all of the centre's electricity has been fed back as heat since November 2024; its impact report gives 65% for 2024, the year of start-up, since heat recovery could not yet run continuously. The company says it offsets 200% of its emissions by funding 2 forestry projects. We found no figure per question.

GreenPT: renewable energy and an indicator for each conversation

GreenPT, a Dutch company, charges for its assistant from €4.50 per month for 15 questions a day, and offers no free plan after a 14-day trial, according to its October 2026 pricing page. It names its hosting providers, Scaleway in France and Verda in Finland, and its models, which range from 24 to 120 billion parameters. Its sustainability page states that all its processing runs in the European Union; its documentation also offers developers a connection endpoint in the United States, the default one remaining in the Union.

GreenPT announces 100% renewable energy; Scaleway's impact report specifies that, for this hosting provider, this means guarantees of origin. GreenPT also cites an efficiency index of 1.25 (the PUE, or power usage effectiveness: the building's electricity divided by that of the servers); in Scaleway's report, that is the figure for a single site, while the average across the provider's sites is 1.37. We do not know which site hosts GreenPT's servers.

GreenPT says it gives each user the energy of their conversations, through a 3-level indicator in the chat and figures in a dashboard. The method published by its co-founder starts from an assumed power of 300 W, calibrated against measurements of graphics processors (GPUs, the computing chips of servers), and counts the tokens produced. We found no public figure per question. GreenPT says it cannot verify that the models were trained with renewable energy.

3Mistral's Le Chat, now Vibe: a published assessment, but no basis for comparison

A life cycle assessment published in July 2025

In July 2025, the French company Mistral AI published the results of the life cycle assessment of its Large 2 model, carried out with the consultancy Carbone 4 and ADEME and reviewed by 2 engineering consultancies: it is the only one we found from a provider. The method also counts the manufacture of the servers. Mistral AI's assistant was then called Le Chat; it became Vibe in May 2026.

For Le Chat, this assessment gives 1.14 gCO2e and 45 mL of water for an answer of 400 tokens (about 320 words), excluding the user's device. We found no complete version of the study and no figure in Wh (watt-hours). A blog post of 22 May 2026 announces a different default model for Le Chat, Mistral Medium 3.5; on Vibe's current pages, we did not find the name of the model that answers. See what this study says about the stages at which AI harms the environment.

What the study cannot tell us

This study cannot show that Mistral AI's assistant emits less, or more, than its competitors. Google published 0.24 Wh and 0.03 gCO2e for a typical written question put to Gemini (the median one when questions are ranked by energy), together with its method. Mistral AI's figure and Google's do not count the same thing: an answer of 400 tokens for the former, a typical question whose length we did not find for the latter.

Electricity is not counted in the same way either. According to Mistral AI's blog post, the study calculates electricity emissions from the location of the servers; we did not find the country or the emission factor there. Google calculates them from its electricity contracts: 94 gCO2e per kWh in 2024, against 345 based on the location of its data centres, according to its own paper.

Mistral AI's study covers the Large 2 model, with data up to January 2025; we found no update since July 2025. Mistral AI proposes that all providers publish 2 indicators: the impact of training a model and that of an answer.

Data in the European Union, but no country found

Mistral AI's help centre states that data is hosted by default in the European Union, or in the United States for developers who choose the connection endpoint there, and refers to a list of subprocessors, which we could not read. We therefore found no country and no hosting provider for the processing of answers. For the European Union as a whole, generating 1 kWh emits 210 gCO2e; among the countries we recorded, the figure runs from 35 in Sweden to 589 in Poland.

In May 2026, Mistral AI announced a 10 megawatt site at Les Ulis, in the Essonne department, with an opening planned for the third quarter of 2026; we found no confirmation of this. When it announced its infrastructure offering Mistral Compute in June 2025, it mentioned the use of low-carbon energy; we found no figure.

4ChatGPT, Claude, Gemini and the others: start with the smaller model

As noted above, Google gives 0.24 Wh for a question put to Gemini, in a methodology paper. For ChatGPT, OpenAI's chief executive stated 0.34 Wh on average on his personal blog; we found no published method. For Claude, we found no figure on the Anthropic pages we read, and none for Lumo (Proton), Perplexity or Qwant on theirs. Our article on the energy of a question put to ChatGPT compares the figures from different sources.

We found no measurement, published by a provider, of the energy gap between its own models. Assuming that price partly follows the amount of computation, prices give an idea: in October 2026, in the price lists we read, the most expensive model costs developers 8 times as much as the cheapest at Google, and 10 times as much at Anthropic.

Our comparison of the ChatGPT, Claude and Gemini models puts them side by side, and the 10 actions ranked by effectiveness quantify what a smaller model changes.

5Running an AI on your own computer saves energy only with a smaller model

How much energy a local model uses

Some software lets you install an open model on your own computer. On a laptop, models with 2 to 8 billion parameters use 0.03 to 0.13 Wh per answer, according to the study already cited for model size, which counts only the processor and the graphics card.

In the same study, a model with 70 billion parameters uses about 2.3 Wh per answer on a server, which is 17 to 70 times more, according to our calculation, for answers of similar length. This ratio combines the effects of model size and of the machine.

A local AI is no more frugal with the same model

The idea that a local AI is necessarily more eco-friendly does not hold when the model is the same. In the study Intelligence per Watt, the chip of a high-end desktop computer, a Mac Studio, needs 1.6 to 2.3 times more energy per request than a data centre card running the same model. The saving appears when the installed model is smaller than the one behind the online service.

According to the same study, if each request were routed to the best of the local models tested, 88.7% of requests sent as a single message would be handled correctly; a single local model on its own does less well.

Who it makes sense for

Running a model locally makes sense for anyone who already owns a recent computer, asks simple questions and wants to keep their data on their own machine. In France, electricity at home has low emissions. Buying a powerful machine for this purpose, however, adds the impact of manufacturing it, which these studies do not count.

6How to choose an eco-friendly AI: 3 questions, and tools for estimating

Do I need a large model?

For summarising or translating, small specialised models use 15 and 35 times less energy respectively than a general-purpose model with 8 billion parameters, according to UNESCO and University College London. If your assistant offers several models, try the lightest one first, and keep reasoning mode for the questions that call for it. Video has a much larger impact: a 5-second video uses as much energy as 300 questions with the generator that serves as the benchmark for this series, one of the most energy-hungry of those measured. See the footprint of an AI-generated video and that of an image.

Where do the servers run?

A service that names its hosting provider and its country gives information that can be checked. Generating 1 kWh in Finland, France or Switzerland emits 7 to 10 times less CO2e than in the United States, according to our calculation from Ember's figures. Among the services we read, Infomaniak and GreenPT name their countries; Mistral AI states the European Union, and for Ecosia we did not find it. See how much electricity data centres consume.

What does the provider publish?

A figure per answer, with its method, can be checked, unlike an adjective. For a target, the date, the scope and the share already achieved must be stated. A company choosing a tool can write these 3 questions into its requirements specification. Our guides explain how to include AI in your Bilan Carbone® (the French carbon accounting method) and measure a company's AI footprint.

Tools for estimating impact

compar:IA, a public comparison tool run by the French Ministry of Culture, rates each model in a class from A to F according to its estimated energy use per million tokens produced, from less than 100 Wh (class A) to more than 10,000 Wh (class F). It uses the method of EcoLogits, which leaves out model training. Our AI footprint calculator gives a company an order of magnitude.

7Key takeaways

• No AI is eco-friendly in itself: model size changes the energy needed for a text about 100-fold.
• Ecosia AI and GreenPT name their models, while Infomaniak and GreenPT name the country of their servers; for Ecosia AI, Euria and GreenPT, we found no public figure per question.
• Mistral AI and Google publish a figure per answer, but on 2 different scopes; for Vibe, we did not find the country of the servers.
• With the same model, a Mac Studio chip needs 1.6 to 2.3 times more energy than a data centre card.

See also: all the impacts of AI, beyond carbon, the water that AI consumes and should you do without AI?.

Further resources

Frequently asked questions

1What is the most eco-friendly AI?

No AI is eco-friendly in itself. We cannot name the most eco-friendly one: we found no figures that can be compared across providers. The footprint of an answer depends first on model size (about 100 times more energy per token for a model with 70 billion parameters than for a small one) and on reasoning mode, then on the country of the servers, which affects the emissions. A small model, without reasoning mode, on servers powered by low-carbon electricity, is the most frugal choice, whichever provider you use.

2Is there a free eco-friendly AI?

Among the services we read, Euria (Infomaniak) and Vibe, formerly Mistral AI's Le Chat, have a free plan in October 2026. For Ecosia AI we found no price displayed; its launch blog post presents Ecosia as free. GreenPT is a paid service, from €4.50 per month after a 14-day trial. What these services publish does not allow their footprints to be compared: among the 4, the only public figure per answer we found is Mistral AI's.

3Is there a Swiss eco-friendly AI, such as Euria?

Among the services we read, Euria is published by the Swiss hosting provider Infomaniak, which presents it as a sovereign and free assistant. According to Infomaniak, its servers are in Switzerland and their heat is fed into Geneva's district heating. The 6,000 homes heated in winter are a capacity targeted by 2028. For 2024, the year the centre started up, the company's impact report puts the share of electricity recovered as heat at 65%. Lumo is published by Proton, another Swiss company: we found no environmental claim on the pages we read.

4Is Mistral's Le Chat, now Vibe, greener than ChatGPT?

We cannot say. In July 2025, Mistral AI published the results of a life cycle assessment: 1.14 gCO2e and 45 mL of water per answer of 400 tokens (about 320 words) from its assistant Le Chat, which became Vibe in May 2026. For ChatGPT, we found only a statement by its chief executive, 0.34 Wh per question, and no published method. The 2 figures differ in unit and in scope, so they do not show which emits less.

5Is running AI locally more eco-friendly?

Not with the same model: according to the Intelligence per Watt study, which was carried out on a Mac Studio, the chip of a desktop computer needs 1.6 to 2.3 times more energy per request than a data centre card. A local AI uses less when the installed model is smaller: 0.03 to 0.13 Wh per answer on a laptop for models with 2 to 8 billion parameters, against about 2.3 Wh for a model with 70 billion parameters on a server, according to another study.

6Does Ecosia have an AI chatbot?

Yes. Ecosia launched Ecosia AI in December 2025; according to its help centre, the chat is in beta and uses the Mistral Small 4 model. We found no price displayed, and the launch blog post presents Ecosia as free. The service generates images but not video, and the AI can be turned off. Ecosia says it produces more clean energy than its AI uses; we found no published figures to support this.

Ask Guillaume your questionA personal reply within 24 working hours

Send my questionor: [email protected]