Artificial intelligence models using reasoning techniques designed to solve complex problems using multiple logical reasoning steps used an average of 100 times more power to respond to written prompts than artificial intelligence models without this capability, a new study found.
The AI Energy Score Project research comes as the technology industry and broader society grapple with the massive power demands of AI applications, which can strain the power grid, drive up energy prices for consumers, and increase environmental pollution from energy production.
The project, led by Sasha Luccioni, a research scientist at Hugging Face, and Boris Gamazaichkov, head of AI sustainability at Salesforce, found that models using inference techniques consumed significantly more power to respond to 1,000 written prompts than models without inference or with inference disabled.
difference in power
The study evaluated 40 open and freely available AI models from companies such as OpenAI, Google, and Microsoft.
One of the companies with the largest difference in power consumption was China's DeepSeek, which gained notoriety in January for its models' ability to deliver high performance with low power consumption.
The R1 model version of DeepSeek required 308,186 watt-hours to complete the task with inference turned on, compared to just 50 watt-hours with the feature turned off.
One of Microsoft's Phi 4 inference models used 9,462 watt-hours to complete the task with inference turned on and 18 watt-hours with inference turned off.
OpenAI's GPT-OSS model saw a reduction in variance, consuming 8,504 Watt-hours to complete the task when inference was set to High and 5,313 Watt-hours when Inference was set to Low.
complex tasks
The researchers ran all the models on the same hardware, using the same prompts, ranging from relatively simple queries to complex math problems.
Project leaders said the study aims to educate users about the power requirements of different models so they can choose the right model for the right task.
Inference models designed for the most complex problems may not be needed for relatively simple queries, they said.
“We need to get smarter about how we use AI,” Luccioni told Bloomberg.
