AI’s Environmental Cost: Comparing Resource Consumption Between SLMs and LLMs Across Queries

Abstract

As artificial intelligence becomes increasingly embedded in daily life, the environmental costs of its deployment remain underexplored. This study investigates the environmental footprint of both large language models (LLMs) and small language models (SLMs); specifically, ChatGPT, Gemini, Deepseek, and Claude, by associating their power draw and water use across queries of varying complexity. Building on evidence that AI services demand substantial resources, this paper asks: how do query complexity and type influence the energy and water consumption of SLMs versus LLMs, and at what threshold of complexity do SLMs become incapable of delivering accurate outputs? To address this, the experimental method categorizes queries into three complexity tiers based on logical steps, conceptual depth, and cognitive skills (recall, evaluation, creation), drawing from the College Board’s question bank of SAT math, reading, and writing problems. Additionally, classic puzzles such as the Tower of Hanoi were selected. Each query was executed three times on the SLM and LLM versions of each commercial AI entity under identical hardware and software configurations. We recorded execution time, model version, and output accuracy. Using the average response time per query, we computed energy consumption and water usage per query. On average, SLMs consumed 60-70% less energy and water than their LLM counterparts, and in subjects such as Math and Reading, had the same level of accuracy as their respective LLMs. However, model performance declined as question difficulty increased, especially in abstract reasoning tasks such as Puzzles, where SLM accuracy dropped considerably. While LLMs were more resource-intensive, they maintained higher accuracy on these challenging queries. SLMs offer a significantly more environmentally sustainable option for simple tasks, but accuracy decreases as complexity increases. A dynamic approach, starting with SLMs and switching to LLMs only when needed, or vice versa, could reduce the environmental cost of AI while maintaining quality. These findings support the potential for context-aware AI deployment strategies that optimize environmental sustainability and accuracy. Future research should aim to quantify this breakpoint more accurately and look at the implementation of automatic query classification systems capable of efficiently switching between models to create more efficient AI models.