Last reviewed:
What is frugal AI? Definition, the AFNOR framework and examples
Frugal AI refers to an AI service whose necessity compared with a less resource-intensive solution has been demonstrated, which applies good practices to reduce its environmental impact, and whose uses have been questioned to stay within planetary boundaries. This is the definition of AFNOR Spec 2314, a French framework published in June 2024 and free to download.
The reference framework is AFNOR Spec 2314, published on 28 June 2024 at the initiative of the French Ministry of Ecological Transition (Ecolab of the General Commission for Sustainable Development), drawn up in six months with some 150 contributors. It sets three conditions for an AI service to be called frugal: the need to use AI rather than a less resource-intensive solution has been demonstrated; producer, supplier and client adopt good practices to reduce impacts; needs and uses have been questioned to stay within planetary boundaries. The voluntary document offers a life-cycle assessment method (hardware manufacturing, training, inference), 31 good-practice sheets, and rules for communicating without greenwashing, for example by specifying the scope of any quantified assessment. It also warns about the rebound effect: a more efficient, hence cheaper, model may be used more and increase total impact. The most common technical levers are well known: choose a small model (SLM) when it is enough, distil a large model into a smaller one, limit unnecessary calls, host in a country with low-carbon electricity. First public case of application: in July 2025 Mistral AI published the life-cycle assessment of Mistral Large 2, carried out with Carbone 4 and ADEME following the AFNOR method: 1.14 g of CO2 equivalent and 45 ml of water for a 400-token response. Frugality differs from efficiency: it starts with the question “do we need AI here?”.
Concrete example
Real case: in July 2025, Mistral AI published the life-cycle assessment of its Mistral Large 2 model, carried out with Carbone 4 and support from ADEME, peer-reviewed by Resilio and Hubblo, following AFNOR's frugal AI methodology. Results: 20.4 kt of CO2 equivalent and 281,000 m³ of water for training and the first months of use (as of January 2025), 1.14 g of CO2 equivalent per 400-token response. The company draws a recommendation from it: choose the model size suited to the need.
Illustrative case: a 120-employee French services SME wants to automatically classify 2,000 customer requests a day. Instead of a large general-purpose model called for every message, it first tests simple rules (keywords), which handle 60% of cases, then a small model hosted in France for the rest. Measured quality is equivalent, the bill and consumption go down, and the approach is documented using the AFNOR Spec 2314 sheets to answer a major customer's CSR questionnaire.
Comparison
| Efficient AI | Frugal AI (AFNOR Spec 2314) | |
|---|---|---|
| Starting question | How to do the same thing with less computation? | Do we need AI here, and at what level? |
| Levers | Smaller model, distillation, quantisation, more efficient hardware | The same, plus questioning the need and limiting uses |
| Measurement | Energy or cost per request | Impacts across the whole life cycle, with scope specified |
| Risk | Rebound effect: more uses, higher total impact | Explicitly addressed by the framework |
| Communication | Efficiency figures | Regulated claims to avoid greenwashing |
FAQ
What is frugal AI?
It is AI used only when it is necessary, designed and run to consume as few resources as possible, and whose uses are questioned to stay within planetary boundaries. The reference definition in France is that of AFNOR Spec 2314 (June 2024).
What is AFNOR Spec 2314?
It is the general framework for frugal AI, published on 28 June 2024 by AFNOR at the initiative of the French Ministry of Ecological Transition. Free and voluntary, it offers a life-cycle assessment method, 31 good-practice sheets and communication rules to avoid greenwashing.
What is an example of frugal AI?
Replacing a large general-purpose model with simple rules and a small model to classify emails, at equal quality. On the vendor side, Mistral AI published in July 2025 the life-cycle assessment of Mistral Large 2 following the AFNOR method, and recommends choosing the model size suited to the need.
What is the difference between frugal AI and sober AI?
The two terms are often synonyms. When distinguished, frugality concerns design and technical operation (model, infrastructure), sobriety concerns uses (how much, for what). AFNOR Spec 2314 covers both.
Is frugal AI mandatory?
No, AFNOR Spec 2314 is voluntary. But it serves as a benchmark in procurement, notably public procurement, and in CSR reporting. At EU level, the AI Act requires providers of general-purpose AI models to document the energy consumption of their models.
How can you make your use of AI more frugal?
First check that AI is necessary, choose the smallest sufficient model (SLM, distilled model), limit calls and response length, reuse results, host in a country with low-carbon electricity, then measure and document the impact.
See also
Further reading
AFNOR Spec 2314, General framework for frugal AI: overview and free download, AFNOR (in French)
Sources
- A framework to measure and reduce the environmental impact of AI (AFNOR Spec 2314), AFNOR Group, 12 July 2024, updated 17 December 2025 (in French). https://www.afnor.org/actualites/intelligence-artificielle/referentiel-reduire-impact-environnemental-ia/
- Publication of the general framework for frugal AI, press release of the French Ministry of Ecological Transition, 28 June 2024 (in French). https://www.ecologie.gouv.fr/presse/publication-du-referentiel-general-lia-frugale-sattaquer-limpact-environnemental-lia
- Our contribution to a global environmental standard for AI, Mistral AI, 22 July 2025 (life-cycle assessment of Mistral Large 2 with Carbone 4 and ADEME). https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai