What a query costs, in rupees and watt-hours

The unit is the token

The chatbot module told you the model’s cost is measured in tokens, not words. This lesson puts prices on it, because knowing the order of magnitude changes how you use the thing and how you read the headlines about it.

A page of English prose (about five hundred words) is roughly six hundred and fifty to seven hundred tokens. Hindi in Devanagari runs two to three times as many tokens per word, for the reason the module on being weaker in your language gave. Keep that page in mind as the unit.

The money

Labs sell access to developers by the million tokens, and they charge more for the tokens a model writes than for the ones it reads, because writing is the sequential, slow part the chip lesson described. In 2025 the prices fell into two bands.

The small models (the ones that answer the free tiers and most business chatbots) cost around ten to fifteen cents of a US dollar per million tokens read and forty to sixty cents per million written. A page in and a page out on one of those is about five hundredths of a US cent: four or five paise.

The frontier models cost roughly twenty times more: a few dollars per million read, ten to fifteen per million written, and considerably more for the models that reason at length before answering. A page in and a page out is about one and a quarter US cents: around a rupee.

A subscription of twenty dollars a month buys, at frontier prices, something like a thousand to two thousand such exchanges, which is why the subscription is the better deal for a heavy user and a poor one for most people, and why the free tier exists at all.

The electricity

Here the numbers are more contested, and it is worth knowing why.

For two years the figure quoted everywhere was about three watt-hours per chatbot query, ten times a web search. It came from a 2023 estimate built on the hardware of the time and the assumption that every query ran on a large model. In 2025 two of the largest labs published their own measurements: a median text prompt of about a quarter of a watt-hour on one and a third on the other, including the idle capacity and cooling around the chip. The fall is real (smaller models and better chips) and so is the caveat: the figures are company-reported, cover a median text query, and exclude training.

To place a third of a watt-hour: a phone battery holds about fifteen watt-hours, so around fifty text queries is one full charge. A ten-watt bulb burns it in two minutes.

Images cost more. A 2023 measurement across many models found generating one image took on the order of a few watt-hours: one image, roughly ten text queries. Video costs more again: minutes of a data-centre chip per clip, tens of watt-hours. And the reasoning models that write thousands of tokens of working before answering cost, per question, ten to a hundred times a plain reply, because the cost is in the tokens written.

Water follows electricity: the same 2025 reports put a text query at a fraction of a millilitre, a few drops, mostly for cooling.

Small each, large together

A third of a watt-hour is nothing. A billion queries a day is roughly three hundred thousand kilowatt-hours a day for text alone, before images, video, reasoning and training. Data centres as a whole drew about one and a half per cent of the world’s electricity in 2024, and the agency that tracks it expects the figure to roughly double by 2030, with AI the fastest-growing part. The final module of this course puts that in its place. For this lesson, the honest position is both halves at once: your query is cheap, and the system is not.

What the training cost adds

Spread the electricity of a frontier training run (tens of gigawatt-hours, from the history module) across the billions of queries the model then answers, and it adds a fraction of a watt-hour to each. Training is a large fixed cost that inference dilutes; within a year of a popular model’s release, serving it has used more power than making it did.

How to use the numbers

Use the small model for the reliable-column tasks; the frontier gains you little on a rewrite and costs twenty times more. Send the page, not the book, when the page is what matters. Do not ask a reasoning model to fix a comma. And when a headline gives a single figure for “the cost of AI”, ask the questions this lesson gave you: per query or per day, text or image, which model, who measured it, and does it include training. The number will usually be true for one of those and reported as if it were true for all of them.