OpenAI’s latest large language model, GPT-5, is facing growing scrutiny from researchers over its potential energy demands and the company’s decision not to disclose official usage figures.

Experts warn that GPT-5’s enhanced capabilities – ranging from building websites to solving PhD-level science problems – may come with a significant environmental cost. Early independent benchmarks indicate the model could consume substantially more electricity per response than previous versions of ChatGPT.

In mid-2023, a simple question such as providing a recipe might have required GPT-4 roughly 2 watt-hours, equivalent to running an incandescent light bulb for two minutes. Research from the University of Rhode Island’s AI lab suggests GPT-5 could use as much as 40 watt-hours for a similar 1,000-token answer, with an average of just over 18 watt-hours – considerably higher than GPT-4o, OpenAI’s previous flagship model.

Nidhal Jegham, a researcher at the lab, said the findings align with expectations for a system believed to be several times larger than GPT-4. Based on ChatGPT’s reported 2.5 billion daily requests, GPT-5’s average consumption rate could equate to the daily electricity use of 1.5 million US homes.

OpenAI has not released model-specific energy data since GPT-3 in 2020, which had 175 billion parameters. While CEO Sam Altman published general resource figures in June – 0.34 watt-hours and 0.000085 gallons of water per query – these were not tied to any specific model and lacked documentation.

Since ChatGPT’s November 2022 launch, which saw the service reach 100 million active users within two months, concerns over the resource demands of generative AI have increased and there’s a case to be made that OpenAI is being intentionally opaque around GPT-5’s energy use.

Mistral AI Study Highlights Broader Environmental Impact

Recent research from Mistral AI adds to the growing body of data on the environmental costs of AI models, covering their intensive training and daily operation.

The French AI developer published a peer-reviewed report on its Mistral Large 2 model, tracking greenhouse gas emissions, water consumption, and materials use. Over less than 18 months, the Large 2 LLM generated 20.4 ktCO2e of greenhouse gases, consumed 281,000 cubic metres of water, and used 660 kg Sb eq of materials. The water usage is comparable to 112 Olympic swimming pools.

Training and inference – the process of running the model – were responsible for 85.5% of all greenhouse gas emissions and 91% of water consumption during this period.

The study also measured the impact of smaller-scale tasks. Generating a single page of text emitted 1.14g CO2e, equivalent to 10 seconds of online streaming in the US or 55 seconds in France. It required 0.05 litres of water, roughly the amount needed to grow a small pink radish, and used materials equivalent to producing a two-euro cent coin (0.2 mg Sb eq).

“These figures reflect the scale of computation involved in Gen AI, requiring numerous GPUs, often in regions with carbon-intensive electricity and sometimes water stress,” the research stated. The analysis also included “upstream emissions” such as those from manufacturing servers.

Recommended reading

• Data Centre Emissions Could Be 600% Worse Than Big Tech Claims
• Report: UK Public Unaware of GenAI’s Environmental Impact
• What is the Carbon Cost of AI?

While individual figures may appear small, usage at scale can cause them to escalate rapidly. The report found a strong correlation between a model’s size and its environmental footprint, noting: “Benchmarks have shown impacts are roughly proportional to model size: a model 10 times bigger will generate impacts one order of magnitude larger than a smaller model for the same amount of generated tokens.”

The findings underline the importance of selecting appropriately sized models for specific use cases, as larger systems tend to have far greater environmental impacts.

Mistral’s research also called for standardised monitoring of environmental impacts across AI models, covering both training and usage, to improve tracking and understanding of their total effect.

While the study focused on greenhouse gas emissions, water consumption, and materials use, it did not detail the effects of extracting or manufacturing those materials, nor the impact of data centres on water pollution and quality.

It did, however, stress the role of geography in mitigating environmental damage. Locating AI training facilities in cooler regions with abundant water resources could reduce the energy and water required for cooling, thereby lowering overall environmental impact.