Skip to content

Short prompt or long prompt: how much does consumption go up?

All comparisons

A long prompt uses 1.5 to 15.1 times more electricity than a short one, and 5.4 times more at the median. That is what Table 4 of the study "How Hungry is AI?" (Jegham et al., arXiv 2505.09598, version of 24 November 2025) gives for the 27 models tested at all three lengths. For GPT-4o, consumption goes from 0.42 Wh to 2.88 Wh, 6.8 times more. The long prompt holds nearly 29 times more tokens (a token is a piece of a word, what the model reads or writes). Pasting a whole document therefore does not cost 29 times the energy, but it still costs a lot more than a question.

Short prompt (GPT-4o, 100 tokens in, 300 out)0.42 Wh
Long prompt (GPT-4o, 10,000 tokens in, 1,500 out)2.88 Wh
Verdict

A long prompt costs 1.5 to 15.1 times more energy than a short one, depending on the model, and 6.8 times for GPT-4o. Cost rises more slowly than length: 29 times more tokens, 5.4 times more energy at the median. But the choice of model weighs far more: on a short prompt, the gap between the most frugal and the hungriest model is 370 times.

Figures updated on October 28, 2026

Mean energy per request by prompt length, eight models from Table 4 of the study
ModelShort (100 tokens in, 300 out)Medium (1,000 and 1,000)Long (10,000 and 1,500)Long ÷ short
GPT-4.10.87 Wh3.16 Wh4.83 Wh5.5 ×
o31.18 Wh5.15 Wh12.22 Wh10.4 ×
GPT-4o (mars 2025)0.42 Wh1.22 Wh2.88 Wh6.8 ×
DeepSeek-R1 (hébergé par DeepSeek)19.25 Wh24.60 Wh29.08 Wh1.5 ×
DeepSeek-V3 (hébergé par DeepSeek)2.78 Wh8.86 Wh13.16 Wh4.7 ×
Claude 3.7 Sonnet0.95 Wh2.99 Wh5.67 Wh6.0 ×
LLaMA-3.1 8B0.05 Wh0.17 Wh0.44 Wh8.5 ×
LLaMA-3.1 70B1.27 Wh4.53 Wh19.18 Wh15.1 ×

Source: Table 4 of Jegham et al., arXiv 2505.09598 (version 6, 24 November 2025). The energy figures are the table's. The "Long ÷ short" column is computed by this site: the study does not publish these ratios. They are modelled estimates, not measurements taken on the servers.

How much more energy does a long prompt use than a short one?

Between 1.5 and 15.1 times more, depending on the model, and 5.4 times at the median across the 27 models that have all three lengths in Table 4. The short prompt is 100 tokens in and 300 out. The long prompt is 10,000 in and 1,500 out, which the study describes as about 7,000 words read and 1,000 words written. For GPT-4o, 0.42 Wh against 2.88 Wh: 6.8 times. For Claude 3.7 Sonnet, 0.95 Wh against 5.67 Wh: 6.0 times. For GPT-4.1, 0.87 Wh against 4.83 Wh: 5.5 times. The table above shows eight models, and the ratios are computed by this site.

Why does 29 times more text not cost 29 times more energy?

Because energy per token falls as the request gets longer. For GPT-4o, a short prompt costs 1.06 Wh per thousand tokens (input and output together), a medium prompt 0.61 Wh, a long prompt 0.25 Wh: four times less per token. The study does not explain it. One possible reading: part of the cost is fixed for each request, and it is spread over more tokens when the request is long. That is a hypothesis, not a result of the study. What to keep is the direction: a long prompt costs more overall, and less per token.

Which models react most to prompt length?

Mostly some LLaMA models. LLaMA-3.1 70B goes from 1.27 Wh to 19.18 Wh: 15.1 times. LLaMA-3.2 Vision 11B, 13.9 times. LLaMA-3.1 405B, 11.3 times. Among reasoning models, o3 goes from 1.18 Wh to 12.22 Wh: 10.4 times. At the other end, DeepSeek-R1 at DeepSeek only goes from 19.25 Wh to 29.08 Wh: 1.5 times, and o4-mini (high) from 3.65 Wh to 7.24 Wh: 2.0 times. Those two already use a lot on a short question, so length adds less in proportion.

Does the choice of model matter more than prompt length?

On this data, yes. On a short prompt, the gap between the most frugal model (LLaMA-3.1 8B, 0.052 Wh) and the hungriest (DeepSeek-R1 at DeepSeek, 19.25 Wh) is 370 times. On a long prompt it is still 65.6 times. For one given model, going from short to long multiplies consumption by 1.5 to 15.1. A gap between models still blends several causes, though: size, reasoning and host. The page "Reasoning model or standard model" on this site isolates the last two.

Does OpenAI's 0.34 Wh figure match a short prompt or a long one?

A short one, according to the study's authors. On 10 June 2025, Sam Altman wrote that an average ChatGPT query uses about 0.34 Wh, without naming the model or the length. The study estimates 0.42 Wh (± 0.13) for a short GPT-4o prompt and finds the gap to be about 19%. A long GPT-4o prompt, at 2.88 Wh, uses 8.5 times that figure. The 0.34 Wh therefore describes an ordinary request, not a whole document pasted in.

Sources

Frequently asked

How many Wh does a long prompt use?+

It depends on the model: from 0.44 Wh (LLaMA-3.1 8B) to 29.08 Wh (DeepSeek-R1 at DeepSeek) across the 27 models in Table 4 of the study "How Hungry is AI?" (version 6, 24 November 2025). GPT-4o: 2.88 Wh. GPT-4.1: 4.83 Wh. Claude 3.7 Sonnet: 5.67 Wh. These are modelled estimates, not measurements taken on the servers.

What do short, medium and long prompts mean?+

They are the three lengths the study tested. Short: 100 tokens in, 300 out. Medium: 1,000 and 1,000. Long: 10,000 and 1,500, which the authors describe as about 7,000 words read and 1,000 written. A two-line question is a short prompt. A contract or a report pasted in whole comes closer to the long prompt.

Why does the site's calculator not distinguish prompt length?+

Because it uses a single value, 0.3 Wh, for an ordinary text prompt: it matches the figures published by Google, Epoch AI and OpenAI, and it is the same order of magnitude as this page's short prompt. A long prompt takes nearly ten times as much for GPT-4o. The "Methodology" page details the assumption.

Are these ratios in the study?+

No. Table 4 gives mean energy in Wh per model and per prompt length, with a standard deviation. The long ÷ short ratios, the median and the energy per token are divisions done by this site on those means. The values hold for these models at the date of the study, not for their successors.

Sending a whole document to an assistant costs more than asking a question, and choosing the right model changes the bill even more. The studio GHIS helps choose and frame these uses before they are deployed.

Go further