The more precisely we attempt to make AI models function, the greater their carbon emissions become — with certain prompts generating up to 50 times more carbon dioxide than others, according to a recent study.

Reasoning models like Anthropic's Claude, OpenAI's o3, and DeepSeek's R1 are advanced large language models (LLMs) that allocate additional time and computing resources to deliver more accurate responses compared to earlier versions.

However, despite achieving impressive results, these models have also demonstrated serious limitations when it comes to solving complex problems. Now, researchers have uncovered another drawback — their substantial carbon footprint. Their findings were published on June 19 in the journal Frontiers in Communication.

"The environmental impact of querying trained LLMs is heavily influenced by their reasoning method, with explicit reasoning methods significantly increasing energy use and carbon emissions," said study lead author Maximilian Dauner, a researcher at Hochschule München University of Applied Sciences in Germany, in a statement. "We observed that reasoning-based models emitted up to 50 times more CO? than models optimized for shorter answers."

When responding to prompts, LLMs process language by dividing it into tokens — segments of words converted into numerical sequences before being processed by neural networks. These networks are refined using training data that determines the likelihood of specific patterns appearing. Based on these probabilities, they generate responses.

To improve accuracy, reasoning models employ a method known as "chain-of-thought." This involves breaking down a complex problem into smaller, logical steps, similar to how humans might solve the same issue.

Related: AI frequently 'hallucinates,' but there's a way to fix it

Sign up for the Live Science daily newsletter nowNonetheless, these models require much more energy than standard LLMs, potentially creating an economic challenge for businesses and users looking to implement them. Despite some existing research on the environmental consequences of expanding AI usage, direct comparisons of the carbon footprints among different models remain limited.

The cost of reasoning

To evaluate the CO? emissions from various models, the researchers behind this new study posed 1,000 questions across multiple topics to 14 LLMs. The models varied in size from 7 billion to 72 billion parameters.

These computations were conducted using the Perun framework — which assesses LLM performance and power consumption — on an NVIDIA A100 GPU. The team then translated energy consumption into CO? emissions, assuming each kilowatt-hour generates 480 grams of CO?.

Their findings indicate that reasoning models generated, on average, 543.5 tokens per question, compared to only 37.7 tokens for concise models. These extra tokens meant increased computational workloads, resulting in higher CO? emissions for the more precise reasoning models.

The most accurate model was the 72 billion parameter Cogito model, which answered 84.9% of the benchmark questions correctly. However, Cogito emitted three times more CO? than similarly sized models designed to produce shorter answers.

"At present, we observe a distinct trade-off between accuracy and sustainability within LLM technologies," Dauner stated. "None of the models that maintained emissions below 500 grams of CO? equivalent [total greenhouse gases released] managed to reach more than 80% accuracy in answering the 1,000 questions correctly."

The above is the detailed content of Advanced AI models generate up to 50 times more CO₂ emissions than more common LLMs when answering the same questions. For more information, please follow other related articles on the PHP Chinese website!