Which AI model should you use for which task?

An AI assistant answers with a model whose size is counted in billions of parameters. The larger the model, the more computation each generated word requires, and therefore the more electricity. Mistral AI sums it up in the life cycle assessment of its Large 2 model: for the same number of tokens, a model 10 times bigger has impacts an order of magnitude higher. For an organisation that sets its teams' tools or builds a service on an API, the choice of model is therefore one of the first settings of its footprint.

The AI Impact Database distinguishes 3 sizes. The small model has about 8 billion active parameters, like Llama 3.1 8B; it consumes about 6 times less per token than the medium model, taken as equivalent to a model of 70 billion, according to the ML.ENERGY benchmark. The medium model is the default one of consumer assistants. The large model, with several hundred billion parameters, consumes about 4 times the medium model, according to the same benchmark. The providers of the large assistants do not publish the size of their models: these classes are reference points.

The table gives the electricity of everyday tasks of an organisation, calculated with the AI Impact Database for each of the 3 sizes, without reasoning. The last column gives the size with which we advise starting trials. No published test covers these tasks as an organisation poses them: we draw this advice from the difficulty of each task (length of the document, number of steps, cost of an error) and from the tests presented below. It should be checked on your own examples.

The gap is the same for almost all rows, because the database applies the same size ratio to text tasks; it is smaller for deep research, where reading web pages does not depend on the model. The volume of the task matters more: a meeting summary written by a large model consumes less than a single deep research run by a small one. For a repeated task, the gap in Wh is multiplied by the number of requests. A customer service handling 100,000 requests a year consumes about 20 kWh with a small model, 120 kWh with the medium model and 490 kWh with a large one, from our calculation on the values in the table.

A large share of an organisation's uses consists of organising information: sorting emails, classifying tickets, spotting a name or an amount in a document, comparing 2 texts. These tasks have a short, well-defined answer. Specialised models, of a few hundred million parameters, have been doing them for years.

Researchers at Hugging Face and Carnegie Mellon measured the same task with the 2 kinds of models. To classify texts by tone, a specialised model emits 0.32 gCO2e for 1,000 queries; general-purpose models of 2.8 to 7 billion parameters emit 2.66 to 4.67 g, that is 8 to 15 times more. For answering questions by extraction, the ratio reaches about 30. The AI Energy Score ranking, which in 2025 measures the energy of recent GPUs alone, finds gaps in the same direction, larger with a bigger general-purpose model: 0.22 Wh for 1,000 classifications with a small specialised model, 28 Wh with a model of 11 billion parameters.

General-purpose models classify slightly better on average, according to the same study, which points out that the 2 kinds of models are hard to compare with the same measures, specialised models having a much narrower choice of answers. The authors conclude that they see no compelling reason to deploy versatile generative models for well-defined tasks, given their consumption. The models they measured date from 2022 and 2023; the gaps between the 2 kinds of models remain the useful information, more than the absolute values.

The tests published by Meta for its Llama 3.1 models show how quality follows size. From the 70 billion parameter model to the 405 billion one, the gain ranges from 1.7 point on a school arithmetic test (GSM-8K) to 8.5 points on a code test (HumanEval). From the 8 billion model to the 70 billion one, it is clearer: 14 points on a general knowledge test (MMLU), 16 points on a mathematics test (MATH) and on doctoral-level questions (GPQA). These tests mostly measure what the model knows and can deduce on its own. When the information is in the request, as for rephrasing, translating or extracting, the gap between sizes is smaller, as the next study shows.

A study by researchers at CNRS and Université Côte d'Azur, published as a preprint in October 2025, looked for 14 tasks for the model with the best balance between quality and size, among those that lose less than 5% quality compared with the best. On average, this model consumes 65.8% less for 3.9% quality lost. For translation, a model of 738 million parameters obtains the same score as a model of 11 billion (32.0 against 32.1); for mathematical reasoning, a task the authors judge less mature, the saving falls to 36%, and the selected model loses about 18% quality there; 4 of the 14 tasks exceed the 5% limit. The hub's analysis of AI decarbonisation works out the effect at world scale.

A team at Berkeley tested a router that sends each request either to GPT-4 or to a smaller model. On a benchmark of varied conversations, 14% of requests sent to GPT-4 were enough to keep 95% of its quality; on knowledge questions, 54% were needed. The share that must go to the large model therefore depends on the type of questions, and a router is trained on requests close to its own. The study's criterion is cost in dollars, not energy.

A reasoning model writes a draft before answering, sometimes thousands of tokens long. The AI Impact Database counts on average 1.92 times the electricity of a direct answer, based on a PEReN (French digital regulation expertise unit) estimate over 22 open models, and about 10 times for deep reasoning. The AI Energy Score benchmark measures 30 times more energy on average for models that reason, and 150 to 700 times more for the same model when its reasoning is switched on. The 2 estimates do not agree: the benchmark measures test questions where the draft multiplies the number of tokens by 300 to 800, which the database retains only for the highest reasoning effort. Depending on the source, reasoning therefore weighs less than size (in the database) or more (in the benchmark).

In the same benchmark, switching on a model's reasoning multiplies its energy by more than moving from Llama 3.1 8B to Llama 3 70B, which multiplies it by about 100. This size ratio depends strongly on measurement conditions: at 64 requests served at once, the ML.ENERGY benchmark, from which the database takes its classes, finds 6.7 times more between the same sizes. The × 24 in the database is therefore the most cautious of the published ratios. Reasoning serves multi-step problems, such as mathematics or code; for writing, translating or classifying, we know of no measurement that shows a quality gain commensurate with its cost. Anthropic states that, for cost and latency, tuning the effort of a single model is often a better lever than switching models. The actions a user can take on reasoning are quantified in the analysis of ways to reduce the impact of your use.

1. For a well-defined, repeated task (classifying, extracting, sorting, detecting), a specialised or small model, chosen on a set of the organisation's own examples.
2. For everyday writing, rephrasing, translation and customer replies, the small model first, without reasoning, and the medium model if trials call for it.
3. For analysing long documents, spreadsheets or contracts, the medium model; the large model when an error has legal, financial or safety consequences.
4. The large model and reasoning for difficult problems: mathematics, complex code; for an agent, the medium model first, the large one if the tasks require it. A router that sends each request to the smallest model able to handle it limits their share.

Providers propose a similar method. Anthropic gives 2 starting points: the smallest model for simple high-volume tasks, moving to a larger model only when needed, or the most capable one for complex reasoning and advanced code. OpenAI advises its most powerful model when cost and latency do not matter, and otherwise to compare results on the same requests and keep the lightest setting that reaches the desired quality level. The choice is made on real examples, because published tests measure exam questions and not an organisation's emails or contracts.