¿Cuánta energía usa un prompt de ChatGPT?

Entre unos 0,3 y 0,4 vatios hora para una pregunta de texto típica. El consejero delegado de OpenAI, Sam Altman, situó la consulta media de ChatGPT en unos 0,34 Wh en junio de 2025, sin publicar el método. Epoch AI estimó de forma independiente unos 0,3 Wh para una consulta típica a GPT-4o, y Jegham et al. estiman 0,42 Wh para un prompt corto de GPT-4o. Eso equivale a unos dos minutos de una bombilla LED de 10 W, o algo más de un segundo de un microondas de 1000 W. Las entradas largas y los modelos de razonamiento gastan mucho más.

¿Un prompt más largo gasta más energía?

Sí, y cuentan tanto el texto que envías como la longitud de la respuesta. CLEER estima que una conversación larga gasta unas nueve o diez veces la energía de una ida y vuelta breve con el mismo modelo. Para prompts únicos, Jegham et al. estiman para GPT-4o 0,42 Wh con un prompt corto (100 tokens de entrada, 300 de salida), 1,2 Wh con uno medio (1000 y 1000) y 2,9 Wh con uno largo (10 000 y 1500). Epoch AI estima que una entrada de 10 000 tokens eleva una consulta a GPT-4o a unos 2,5 Wh, y una de 100 000 tokens a casi 40 Wh.

¿Por qué no aparecen GPT-6 ni Claude Fable 5.1?

Porque todavía no hay ninguna cifra publicada que los cubra. Las empresas de IA no informan del consumo de energía por modelo: Google y OpenAI solo han publicado medias de sus aplicaciones, y Anthropic nada en absoluto. Los modelos actuales de esta página usan las estimaciones independientes de CLEER, y su última instantánea, de septiembre de 2026, es anterior a GPT-6 y Claude Fable 5.1. En lugar de adivinar, esta calculadora los deja fuera hasta que se publique una cifra fiable.

¿Un prompt de IA gasta más energía que una búsqueda en Google?

Un prompt corto típico está ahora en el mismo orden. Google dijo en 2009 que una búsqueda usaba unos 0,3 Wh, y las cifras actuales para un prompt corto de chat van de unos 0,24 Wh (la mediana de Gemini de Google) a unos 0,4 Wh. Pero un documento largo o un modelo de razonamiento pueden gastar de diez a cien veces más: Jegham et al. estiman más de 12 Wh para un prompt largo de o3. La cifra de búsqueda de Google es de 2009, así que toma la comparación como aproximada.

¿Cuánta agua usa un prompt de IA?

Las cifras publicadas van de menos de medio mililitro a decenas de mililitros, sobre todo porque cuentan cosas distintas. Google sitúa la mediana de un prompt de Gemini en 0,26 mL, unas cinco gotas, contando el agua que consumen sus centros de datos. Sam Altman dio 0,000085 galones, unos 0,32 mL, para una consulta media de ChatGPT. El análisis de ciclo de vida de Mistral sitúa una respuesta de 400 tokens de su asistente Le Chat en 45 mL, porque es una cifra de ciclo de vida completo que también cuenta impactos previos como la fabricación de los servidores.

¿Por qué las estimaciones publicadas difieren tanto?

Miden cosas distintas en hardware distinto. Los 0,24 Wh de Google incluyen la capacidad inactiva, las CPU y la memoria de los servidores y el consumo adicional del centro de datos; si solo se cuentan los chips de IA, salen 0,10 Wh. Jegham et al. incluyen el consumo adicional del centro de datos pero dejan fuera las máquinas inactivas. El mismo modelo también puede variar mucho según dónde se ejecute: Jegham et al. estiman que DeepSeek-R1 gasta unas ocho veces más energía por prompt corto en los servidores de DeepSeek que en Microsoft Azure. Toma cada cifra de esta página como una orientación de escala, no como una medición exacta.

AI Energy Calculator

Pick a type of conversation and a model to see how much electricity it takes, then compare it with a light bulb, a microwave, phone charges, coal, barrels of oil and more — for one conversation or billions.

Every figure is sourced and labelled

No AI company publishes energy use for each of its models, so current models use CLEER's independent estimates (Sustainable AI Group, September 2026), built from thousands of measured runs of open models. Older models and the providers' own published figures are here too. GPT-6 and Claude Fable 5.1 are not covered by any published figure yet, so they are left out rather than guessed.

1. Choose a type of conversation

Quick question One exchange “What's the difference between affect and effect?” A definition, a short explanation or a one-off task, answered in a single exchange. Short back-and-forth A few exchanges “Plan a three-day trip to Lisbon.” Then: “Make day two cheaper.” Then: “Add a rainy-day option.” A question, an answer, and a follow-up or two to refine or clarify it. Extended conversation Many turns, with reasoning “Help me decide whether to move into data science,” followed by an hour of questions about courses, pay and risks. Many turns exploring a topic in depth, with the model reasoning at length before each reply. Vibe coding session An hour or two with a coding agent “Build me a habit-tracker web app with streaks and dark mode,” then letting the agent write, run and fix the code. The assistant reads files, makes changes and checks its own work across a project.

2. Choose a model

3. Conversations per person

4. Team size (people)

Energy per conversation · 0 Wh

Total energy 0 Wh

What That Much Energy Compares To

A 10 W LED bulb stays lit for — 10 W is an assumed bulb wattage

A 1,000 W microwave runs for — 1,000 W is an assumed microwave wattage

Full smartphone charges — 19 Wh per charge (US EPA)

Video streaming for — 77 Wh per hour (IEA, 2019 data)

Google searches — 0.3 Wh per search (Google, 2009)

Litres of water boiled — 93 Wh to heat 1 litre from 20 °C to boiling

Coal a power station burns to make it — 1.14 lb per kWh generated (US EIA, 2022)

Oil a power station burns to make it — 0.08 US gallons per kWh generated (US EIA, 2022)

CO₂ emitted — World-average electricity, 473 g per kWh (Ember, 2024)

Same CO₂ as driving a petrol car — 393 g of CO₂ per mile (US EPA)

Water used —

Current Models for This Conversation, Lowest Energy First

These are CLEER's estimates. They share one method, so they can be ranked against each other; ± shows one standard deviation. Older estimates and provider figures use different methods, so they are in the model list but not in this table.

Model	Source of figure	Per conversation	Total	10 W LED bulb lit for

Figures last checked: October 26, 2026

Energy per prompt depends heavily on the hardware, the data centre and how busy the servers are. Treat every figure as a guide to scale, not a meter reading.

How the Calculator Works

Each model has an energy figure per conversation, in watt-hours (Wh). The calculator multiplies it by the number of conversations each person has and by the size of the team, then divides the total by what each comparison uses:

total energy  = energy per conversation × conversations per person × team size
bulb time     = total energy ÷ 10 W
phone charges = total energy ÷ 19 Wh
coal burned   = total energy in kWh × 1.14 lb
CO₂           = total energy in kWh × 473 g

For example, the average ChatGPT query at 0.34 Wh, sent 20 times a day for a year, comes to 7,300 prompts and about 2.5 kWh: roughly 130 phone charges, 10 days of a 10 W LED bulb, or 1.2 kg of CO₂ on a world-average grid — about the same CO₂ as driving a petrol car 3 miles.

Where the Figures Come From

Each figure is labelled with how it was produced:

Independent estimate (CLEER). For current closed models, the Sustainable AI Group's CLEER project (September 2026) ran more than 2,400 tests on over 20 open models, measuring their power draw on real hardware. It matches each closed model to the open models it most resembles in size, capability and speed, and projects its energy from theirs. The figures are total data-centre energy — chips, host servers, idle capacity and cooling — for three kinds of chat conversation, each with a standard deviation and a confidence score.
Measured (open model). Open models that CLEER ran itself, measured directly and scaled to total data-centre energy in the same way.
Measured by provider. Google published a measured median for Gemini Apps text prompts in August 2025: 0.24 Wh, 0.03 g of CO₂e and 0.26 mL of water, from May 2025 data, with its method in an accompanying paper.
Stated by provider. OpenAI's chief executive gave an average of about 0.34 Wh per ChatGPT query in June 2025, with no method or scope published.
Earlier estimate. Jegham et al.'s 2025 benchmark, "How Hungry is AI?", estimates energy for older models from public API speed data, for single prompts of fixed length: 100 tokens in and 300 out, 1,000 in and 1,000 out, and 10,000 in and 1,500 out. This page uses them as the nearest match to each conversation type, and includes the paper's GPT-5 figures at each reasoning setting.
Research median. Microsoft Research's 2026 paper in Joule models a median of 0.31 Wh per query for frontier-scale models (over 200 billion parameters) on H100 servers, rising to 3.91 Wh for long reasoning queries. It is not specific to any one model.

Anthropic publishes no per-prompt energy figure, and neither OpenAI nor Google publishes one per model. GPT-6 and Claude Fable 5.1 are newer than CLEER's September 2026 snapshot, so no published figure covers them yet.

Why Coal and Oil Are Counted at the Power Station

A barrel of oil holds far more energy as heat than the electricity it can make, because power stations lose much of a fuel's heat in turning it into electricity. So the fair comparison is the fuel burned per kilowatt-hour actually generated. The US Energy Information Administration publishes it: in 2022, 1.14 pounds of coal or 0.08 US gallons of petroleum per kWh, which is about 540 kWh of electricity per 42-gallon barrel.

The coal and oil figures here are what a coal- or oil-fired power station would burn to supply the same electricity. AI data centres run on a mix of sources, so read them as a sense of scale, not as fuel that was actually burned.

Small Per Prompt, Large at Scale

One prompt's energy is tiny: the oil needed to generate 0.34 Wh is about a tenth of a millilitre. But OpenAI told Axios in July 2025 that ChatGPT handles more than 2.5 billion prompts a day. At 0.34 Wh each, that is about 850 MWh a day — the electricity from roughly 1,600 barrels of oil or 440 tonnes of coal burned in power stations, or enough to keep a 10 W LED bulb lit for about 9,700 years. Use the "ChatGPT's daily prompts" button above, with the ChatGPT average selected, to see the full comparison.

Coding agents multiply this. Across the models CLEER covers, the median vibe coding session uses about 27 times the energy of a short back-and-forth with the same model, because the agent reads files, runs tools and checks its own work over many steps. The vibe coding option above is CLEER's typical agentic session: an hour or two of an assistant working across a project.

How to Use Less Energy with AI

Pick a smaller model. For a short back-and-forth, CLEER estimates Gemini 3.5 Flash-Lite at 0.036 Wh and Claude Fable 5 at 2.8 Wh — nearly 80 times more. Classification, extraction and short rewrites rarely need a large model.
Save reasoning for hard problems. Reasoning models write hidden thinking tokens before they answer. Use a standard model, or a low reasoning setting, when the question is simple.
Send less context. Input length drives energy as well as cost. Send the relevant passage rather than the whole document, and start a new chat instead of carrying a long history.
Ask once, clearly. Every regenerated answer costs a full prompt's energy again. A precise first prompt saves energy, time and money.

Related Tools

Not sure which prompt size matches yours? Count its tokens, and see what it costs, before you choose. Measure it with the LLM Token Counter.

References

Frequently Asked Questions

How much energy does one ChatGPT prompt use?

About 0.3 to 0.4 watt-hours for a typical text question. OpenAI's chief executive, Sam Altman, put the average ChatGPT query at about 0.34 Wh in June 2025, without publishing a method. Epoch AI independently estimated about 0.3 Wh for a typical GPT-4o query, and Jegham et al. estimate 0.42 Wh for a short GPT-4o prompt. That is roughly two minutes of a 10 W LED bulb, or just over a second of a 1,000 W microwave. Long inputs and reasoning models use much more.

Does a longer prompt use more energy?

Yes, and both the text you send and the length of the reply count. CLEER estimates that an extended conversation uses about nine to ten times the energy of a short back-and-forth on the same model. For single prompts, Jegham et al. estimate GPT-4o at 0.42 Wh for a short prompt (100 tokens in, 300 out), 1.2 Wh for a medium one (1,000 in, 1,000 out) and 2.9 Wh for a long one (10,000 in, 1,500 out). Epoch AI estimates that a 10,000-token input raises a GPT-4o query to around 2.5 Wh, and a 100,000-token input to almost 40 Wh.

Why aren't GPT-6 and Claude Fable 5.1 listed?

Because no published figure covers them yet. AI companies do not report energy use per model: Google and OpenAI have published only app-wide averages, and Anthropic nothing at all. The current models on this page use CLEER's independent estimates, and its latest snapshot, from September 2026, predates GPT-6 and Claude Fable 5.1. Rather than guess, this calculator leaves them out until a credible figure is published.

Does an AI prompt use more energy than a Google search?

A typical short prompt is now in the same range. Google said in 2009 that a search used about 0.3 Wh, and current figures for a short chat prompt run from about 0.24 Wh (Google's median Gemini prompt) to about 0.4 Wh. But a long document or a reasoning model can take ten to a hundred times more: Jegham et al. estimate over 12 Wh for a long o3 prompt. Google's search figure is from 2009, so treat the comparison as rough.

How much water does an AI prompt use?

Published figures range from under half a millilitre to tens of millilitres, mostly because they count different things. Google puts its median Gemini prompt at 0.26 mL, about five drops, counting the water its data centres consume. Sam Altman gave 0.000085 gallons, about 0.32 mL, for an average ChatGPT query. Mistral's lifecycle study puts a 400-token response from its Le Chat assistant at 45 mL, because it is a full lifecycle figure that also counts upstream impacts such as manufacturing the servers.

Why do published estimates disagree so much?

They measure different things on different hardware. Google's 0.24 Wh includes idle capacity, host CPUs and memory, and data-centre overhead; counting only the AI chips gives 0.10 Wh. Jegham et al. include data-centre overhead but leave out idle machines. The same model can also differ widely by host: Jegham et al. estimate DeepSeek-R1 at about eight times more energy per short prompt on DeepSeek's own servers than on Microsoft Azure. Read every figure here as a guide to scale, not an exact measurement.