A new study has found that large language models (LLMs) designed to deliver more detailed, step-by-step answers – known as “reasoning models” – generate significantly more carbon emissions than simpler, faster AI systems. In some cases, these models emit up to 50 times more CO₂ per query, raising serious concerns about the environmental footprint of artificial intelligence.

Published in Frontiers in Communication, the research, titled Energy costs of communicating with AI compared 14 different LLMs by tasking them with 1,000 benchmark questions across subjects like history, algebra, and philosophy. Researchers measured both the accuracy of their answers and the volume of “tokens” generated per response – tokens being the units of data used by AI to process and produce text.

The findings reveal a striking environmental cost associated with more advanced AI outputs. On average, reasoning models generated 543.5 tokens per question, while concise models produced just 37.7.

Since each token consumes energy, this difference translated directly into higher emissions. In fact, the most accurate reasoning model tested – Cogito – produced three times as much CO₂ as similarly sized models that delivered briefer answers.

“The environmental impact of questioning trained LLMs is strongly determined by their reasoning approach,” said lead author Maximilian Dauner, a researcher at Hochschule München University of Applied Sciences.

“Explicit reasoning processes significantly drive up energy consumption and carbon emissions.”

A Hidden Trade-Off: Accuracy vs Sustainability

At the heart of the findings is a trade-off that’s becoming increasingly difficult to ignore: greater accuracy and nuance come at the cost of sustainability.

“None of the models that kept emissions below 500 grams of CO₂ equivalent achieved higher than 80% accuracy,” said Dauner. The most precise model tested, Cogito, reached 84.9% – but at far greater environmental cost.

Reasoning models are typically used in more open-ended or abstract tasks that require logical progression and deeper analysis. These include questions related to philosophy, advanced mathematics, or complex ethical scenarios. In contrast, straightforward factual queries -like those about high school-level history – were processed with far fewer tokens and produced significantly lower emissions.

Even among models with similar sizes, emissions varied dramatically depending on how they processed information.

DeepSeek R1, a 70-billion-parameter model, would emit as much CO₂ as a round-trip flight from London to New York if asked to answer 600,000 questions. By contrast, the Qwen 2.5 model, with 72 billion parameters, could handle three times as many questions before hitting the same emissions level – around 1.9 million queries in total.

While parameters (the internal variables AI models use to “learn”) are a factor, the study suggests emissions are more tightly linked to how models generate answers – especially whether they engage in chain-of-thought reasoning or prioritise brevity and speed.

Growing Demand, Growing Footprint

The research arrives amid a broader reckoning over AI’s mounting energy demands. Major tech companies including Apple, OpenAI, Oracle, and SoftBank are collectively investing hundreds of billions of dollars in AI-dedicated data centers. These facilities house the hardware required to run LLMs and are becoming major energy consumers in their own right.

According to the Electric Power Research Institute, AI-driven data centers could account for up to 9.1% of total U.S. electricity usage by 2030 – more than double their current share. Each AI query can be ten times more energy-intensive than a standard Google search, the International Energy Agency has warned.

Beyond this, research by Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam’s Institute for Environmental Studies, AI may account for nearly half of global data centre electricity usage by the end of 2025 – up from an estimated 20% today.

The study’s findings suggest that without changes to model architecture or usage habits, the rise of reasoning-intensive AI tools could substantially accelerate that growth.

Recommended reading

• £4M Boost for Scottish Clean Energy Projects
• Will Great British Energy Be Headquartered in Aberdeen?
• UK Gov Planning Bill to Fast-track Energy Projects

The researchers hope their findings will encourage more informed and sustainable use of LLMs.

“Users can significantly reduce emissions by prompting AI to generate concise answers or limiting the use of high-capacity models to tasks that genuinely require that power,” said Dauner.

Knowing the carbon cost of everyday AI interactions – whether asking for a recipe or generating a fictional short story – could help curb unnecessary usage.

“If users know the exact CO₂ cost of their AI-generated outputs,” Dauner added, “they might be more selective and thoughtful about when and how they use these technologies.”