OpenAI will not disclose GPT-5’s energy use. It could be higher than past models | OpenAI

In in mid -2023, if a user asked for the Openai Chatppt for an artichoke pasta recipe or instructions on how to make a ritual offering at the old cananate divinity Moloch, his answer could have taken – very closely – 2 Watthers, or almost as much electricity as a incandescent bulb consumes in 2 minutes.

OPENAI published a model on Thursday that underlies the popular chatbot-GPT-5. Ask this version of the AI for an artichoke recipe, and the same amount of text related to pasta could take several times – or even 20 times – this amount of energy, according to experts.

As its deployment of the GPT-5, the company has stressed the revolutionary capacities of the model: its ability to create websites, to answer science questions at doctorate and reason through difficult problems.

But experts who have spent the past few years working to compare the energy and the use of the resources of AI models say that these new powers cost at a cost: a GPT-5 response can take a much greater amount of energy than a response from the previous versions of Chatgpt.

OPENAI, like most of its competitors, has not published any official information on the use of electricity of its models since GPT-3, which was released in 2020. Sam Altman, its CEO, threw some figures on the consumption of chatgpt resources on its blog in June. However, these figures, 0.34 watthers and 0.000085 water gallons per request, do not refer to a specific model and have no support documentation.

“A more complex model like GPT-5 consumes more power both during training and during inference. It is also targeted at a long reflection … I can say safely that it will consume much more power than the GPT-4, “said Rakesh Kumar, professor at the University of Illinois, currently working on the consumption of calculation energy and AI models.

The day of the GPT-5 was published, researchers from the IA laboratory of the University of Rhode Island revealed that the model can use up to 40 wattheures of electricity to generate an average response of around 1000 tokens, which are the constituent elements of the text for an IA model and are approximately equivalent to words.

A dashboard that they display on Friday indicates that the average energy consumption of GPT-5 for an average response is just over 18 wattheures, a figure which is higher than all the other models they compare with the exception of the O3 reasoning model, published in April and R1, produced by the Chinese Deepseek company.

It is “much more energy than GPT-4O”, the previous model of Openai, said Nidhal Jegham, group researcher.

Eighteen Watthers would correspond to the combustion of this incandescent bulb for 18 minutes. Given the recent reports that Chatgpt manages requests of 2.5 billion billions of dollars a day, the total consumption of GPT-5 could reach daily electricity demand of 1.5 million American houses.

As large as these figures are, researchers on the ground say that they align with their large expectations for the energy consumption of GPT-5, since the GPT-5 would be several times larger than the previous models of Openai. OPENAI has not published the number of parameters – which determine the size of a model – for any of its models since GPT -3, which had 175 billion parameters.

Disclosure this summer of French Society of the Mistral this summer finds a “strong correlation” between the size of a model and its energy consumption, based on the study of Mistral on its internal systems.

“Depending on the size of the model, the quantity of resources [used by GPT-5] should be superior orders than that of GPT-3, “said Shaolei Ren, professor at the University of California in Riverside who studies the imprint of AI resources.

Benchmarking AI Use of power

The GPT-4 was widely considered 10 times the size of GPT-3. Jegham, Kumar, Ren and others say that the GPT-5 is probably much larger than GPT-4.

The main companies of AI as Openai believe that extremely important models may be necessary to reach act, that is to say an AI system capable of making human jobs. Altman strongly pleaded for this point of view, writing in February: “It seems that you can spend arbitrary sums and obtain continuous and predictable gains”, although he declared that the GPT-5 had not exceeded human intelligence.

In his comparative analysis study in July, which examined energy consumption, the use of water and carbon emissions for the bot The cat of Mistral, the startup found a relationship one to one between the size of a model and its consumption of resources, writing: “A model 10 times larger will generate impacts.

Jegham, Kumar and Ren said that even if the GPT-5 scale is important, there are probably other factors that come into play to determine its consumption of resources. GPT-5 is deployed on more efficient equipment than certain previous models. The GPT-5 seems to use an architecture “mixture of experts”, which means that it is rationalized so that all its parameters are not activated when they respond to a request, a construction which will probably reduce its energy consumption.

On the other hand, the GPT -5 is also a model of reasoning, and works in the video and the images as well as the text, which probably makes its energy footprint much greater than text operations only, say Ren and Kumar – especially since the mode of reasoning means that the model will be calculated longer before responding to a request.

“If you use the mode of reasoning, the quantity of resources you spend to get the same answer will probably be several times higher, from five to 10,” said Ren.

Hidden information

In order to calculate the consumption of resources of an AI model, the group of the University of Rhode Island has multiplied the average time that the model takes to respond to a request – whether for a pasta recipe or an offering to Moloch – by the average power draw of the model during its operation.

The estimate of the draw for a model was “a lot of work,” said Abdeltawab Hendawi, a data science professor at the University of Rhode Island. The group has struggled to find information on how different models is deployed in data centers. Their final article contains estimates for which the chips are used for a given model, and how different requests are offset between different chips in a data center.

Altman’s June blog confirmed their conclusions. The figure he gave for Chatgpt energy consumption per request, 0.34 Watthers per request, corresponds closely to what the group has found for GPT-4O.

Hendawi, Jegham and other members of their group declared that their conclusions underlined the need for more transparency in AI companies when they publish ever heavier models.

“It is more critical than ever to treat the true environmental cost of AI,” said Marwan Abdelatti, professor in Uri. “We call Openai and other developers to use this moment to engage in total transparency by publicly revealing the environmental impact of GPT-5.”