Do we really know how much electricity ChatGPT uses?

In June 2025, Sam Altman, the head of OpenAI, writes in a blog post that an average request to ChatGPT consumes "about 0.34 watt-hour", and about 0.3 mL of water. The post says neither what is counted, nor on which models, nor how. It is, to date, the only value given by OpenAI or its head; its training page on the environment, updated in June 2026, refers to the Epoch AI estimate, about 0.3 Wh, and states that its own measurements are not yet published.

For a short written question, the order of magnitude of 2025 is nevertheless known. The Google measurement on Gemini, the Epoch AI estimate for GPT-4o and the OpenAI statement converge around 0.3 Wh, the value used by the AI Impact Database, and detailed by our article on the consumption of a request↗. Since August 2025, ChatGPT has run on the GPT-5 family, for which OpenAI has published no value; an independent estimate gives 2.33 Wh for a request of average length to GPT-5 at minimal reasoning effort. This order of magnitude also says nothing about the total consumption of an assistant or that of heavier uses.

For 8 major assistants (ChatGPT, Gemini, Copilot, Meta AI, Claude, Grok, DeepSeek and Vibe, formerly Le Chat), we looked for what their publisher has itself published on the electricity, water or carbon of a request, up to 13 November 2026.

The 2 values published with a method cannot be compared with one another. Google counts the electricity of the median request and emissions calculated with its renewable electricity contracts; Mistral AI counts the life cycle of a response of 400 tokens, including the manufacture of the servers. Our analysis of diverging estimates sets out these differences in scope.

This finding matches that of a team led by Sasha Luccioni of Hugging Face. On OpenRouter, a platform that gives access to many models through an API, 84% of the tokens processed in May 2025 by the 20 most-used models went through models with no published environmental information. For notable AI models, the team observes that such disclosures, which peaked in 2022, declined with the commercial models launched after ChatGPT. The 2 values published with a method date from 2025 and have not been updated for current models.

To go from a value per question to the consumption of an assistant, it would have to be multiplied by a number of requests that publishers sometimes announce, and by a mix of uses that they do not publish. We found, for none of the 8 publishers, the total electricity of its assistant.

The uses that weigh the most have no value published by their publisher: image and video generation, reasoning mode, agents. Yet these are what make a year of use range from a few hundred grams to more than 100 kgCO2e. Finally, for a consumer user, we found none of the 8 assistants that states in which country its request is processed. Yet the country of the servers changes the CO2e of the same electricity by a factor of about 21, according to the AI Impact Database.

Since 2 August 2025, the EU AI Act requires providers of general-purpose AI models to keep technical documentation that includes the known or estimated energy consumption of the model, in the part devoted to its development, and so to its training rather than its use. Models placed on the market before that date have until 2 August 2027. This documentation is handed over on request to the Commission's AI Office and to national authorities; it is not published. The code of practice of July 2025, which publishers sign voluntarily, merely encourages them to consider what they could make public.

In France, ADEME notes in its July 2026 opinion that the lack of transparency of the main players makes it harder to measure impacts. It calls for reliable and comparable figures, a multi-criteria life cycle assessment, and emissions calculated with the grid electricity where the servers are connected, rather than with power purchase agreements.

The strongest objection is commercial secrecy. Publishing the electricity of a request, the number of requests and the location of the servers would give competitors the size of its models, its efficiency and its customer base, in a market where this information is valuable. The figures also change very fast: Google announces a 33-fold drop in the energy of the median Gemini request in one year, from May 2024 to May 2025.

It does not convince us, for 3 reasons. Google has published its method without publishing its volumes, and OpenAI already announces its own: 2.6 billion requests per day in July 2025, a figure that ADEME repeats. Publishing a method or a total therefore gives away no secret that these companies do not already give away. The speed of change argues for dated figures reviewed every year, rather than for no figures. And the AI Office can already request the energy of models: the question is what the public can see of it.