IN-In mid-2023, when users asked Artichoke Pasta recipes or instructions on how to serve rituals from the ancient Canaanite god Moloch, the response was very much powered by an incandescent bulb in about 2 watt hours or as much as two minutes.
Openai released a model on Thursday that supports the popular chatbot GPT-5. When you ask the AI version about Artichoke's recipes, experts say that the same amount of pasta-related text could take that amount of energy several times (20 times).
With the deployment of the GPT-5, the company created a breakthrough feature for the model, namely website, answering PHD-level science questions, highlighting the reason through difficult questions.
However, experts who have worked for the past few years to benchmark the energy and resource usage of AI models say these new forces are at the expense. Responses from GPT-5 can take significantly more energy than responses from previous versions of ChatGPT.
Openai, like most of its competitors, has not released official information on the model's power usage since the GPT-3, which was announced in 2020. CEO Sam Altman threw out some numbers on ChatGPT's resource consumption this June. However, these numbers, 0.34 watt hours, water per query of 0.000085 gallons, do not refer to any particular model and there are no support documentation.
“More complex models like the GPT-5 consume more power during training and inference. They also target long-term thinking. They can safely say they consume much more power than the GPT-4.”
Researchers at the University of Rhode Island AI Lab found that the model could use up to 40 watts of electricity to generate a medium-length response of about 1,000 tokens. This is a component of the text of the AI model and is roughly equivalent to words.
A dashboard installed on Friday shows that the average energy consumption of the GPT-5 for medium-length responses is just over 18 watts. This is a higher diagram than all other models except for the OpenAI O3 inference model released in April, and the R1 created by Chinese AI company Deepseek.
This is “significantly more energy than Openai's previous model, GPT-4O,” says Nidhal Jegham, a researcher in the group.
One watt of 18 watt hours corresponds to burning that incandescent light bulb for 18 minutes. Given the recent reports that ChATGPT handles 2.5 billion requests per day, total GPT-5 consumption could reach daily electricity demand in the US home at a 1.5m.
As much as these numbers, researchers in this field say they are in line with the broad expectations of GPT-5's energy consumption, given that the GPT-5 is considered several times larger than Openai's previous model. OpenAI has not released a parameter count for any of the models since GPT-3 with 175bn parameters to determine the size of the model.
This summer, a disclosure from French AI company Mistral found a “strong correlation” between model size and energy consumption, based on research into Mistral's internal systems.
“Amount of resources based on model size [used by GPT-5] “We're a great fan of AI,” said Shaolei Ren, a professor at the University of California, Riverside, who studies AI resource footprints.
AI power usage benchmark
The GPT-4 was widely believed to be 10 times the size of the GPT-3. Jegham, Kumar and Ren say that GPT-5 is likely to be significantly larger than GPT-4.
Large AI companies like Openai believe that they need a very large model to achieve AGI, an AI system that can do human work. Altman made a strong point about this view, writing in February. “It seems like you can spend arbitrary amounts and get a continuous, predictable profit.”
After the newsletter promotion
In a July benchmark survey examining Mistral's LE chatbot's power consumption, water usage and carbon emissions, startups discovered a one-to-one relationship between model size and resource consumption, writing:
Jegham, Kumar and Ren said that while GPT-5 scale is important, there are likely other factors that can affect resource consumption. The GPT-5 is deployed on more efficient hardware than its previous models. GPT-5 appears to use the “Experts” architecture. This means that not all parameters are activated when responding to a query.
Meanwhile, the GPT-5 is also an inference model, which works with video, images and text. This is likely to result in a much larger energy footprint than text-only manipulation, Ren and Kumar say.
“With inference mode, the amount of resources you spend to get the same answer can be five to ten times more,” Ren said.
Hidden information
To calculate the resource consumption of AI models, a group at the University of Rhode Island multiplied the average time the model needs to respond to queries – by the average power draw of the model in operation, such as pasta recipes or offerings to Morocco.
Abdeltawab Henderwi, a professor of data science at the University of Rhode Island, said estimating the model's power draw was “a lot of work.” This group struggled to find information about how different models were deployed within the data center. Their final paper includes estimates where chips are used for a particular model, and how different queries are split between different chips within a data center.
Altman's June blog post confirmed their findings. The numbers he gave for energy consumption per query on CHATGPT, 0.34 watt-hour business hours per query, are closely matched to what the group found on GPT-4O.
Others in the group, including Hendawi, Jegham, and others said their findings underscore the need for more transparency when AI companies release models that have never been seen before.
“It's more important than ever to address the true environmental costs of AI,” says Marwan Abdelatti, professor at URI. “We are calling on Openai and other developers to use this moment to commit to full transparency by exposing the environmental impact of GPT-5.”
