Why do reasoning modes and agents use so much more energy?

The most frequently cited figures on the energy of AI relate to a short query to an assistant, of the order of 0.3 Wh. At the end of 2025, however, reasoning models handled more than half of the tokens passing through the OpenRouter platform, which mainly serves developers, and agents chain model calls on their own. This explainer follows the text the model writes and rereads in each mode, because the energy depends on it; the analysis of the change over time follows the growth of these uses.

A language model splits text into tokens, fragments of words. It writes its response one token after another, and each token written requires a pass through the whole model, on GPUs (the explainer on GPUs explains why). Before writing, it reads all the text it is given: the question, the previous conversation, the attached documents. The energy therefore has 2 parts, one that follows the length of the text written, the other that of the text read.

Epoch AI estimates 0.3 Wh a query to GPT-4o that writes 500 tokens, and Microsoft estimates the energy of a median query in an optimised deployment at 0.31 Wh. The text read weighs little as long as it stays short. With 10,000 input tokens, about 20 pages, Epoch AI's estimate rises to 2.4 Wh; with 100,000 tokens, to 40 Wh, because the computation for reading grows faster than the length of the text when it becomes long.

In reasoning mode, the model first writes an intermediate text, in which it breaks the question down and checks its steps, then its response. The user usually sees only a summary of it, or nothing, but each token of this draft costs as much as a token of the response. Hugging Face measured 3 open models with and without reasoning: they write 300 to 800 times more tokens and consume 154 to 697 times more energy with reasoning switched on, depending on the model. On average, the reasoning models in its ranking consume 30 times more than the models without reasoning, a comparison between families of models rather than between 2 settings.

The gap depends on the length of the draft, and therefore on the task and the setting. PEReN, for Arcep, measured 22 open models: switching on reasoning adds 92% of energy on average, with a 95% interval from -21% to +348%, and 849% on average on programming exercises, against 41% on general knowledge questions. For GPT-5, the estimate by Jegham and co-authors ranges from 2.33 Wh with minimal reasoning effort to 17.15 Wh with high effort, for a question of average length, which is 7 times more. On a simple text task, the IEA counts about 2 times the electricity of a model of comparable size.

Uses

Reasoning mode multiplies energy use

Energy with reasoning, relative to the same request without, as a multiple · Measurements from 2025 and 2026

PEReN, average22 open models× 1.9

PEReN, codemaximum, one task× 9.5

Hugging Face, average2 model groups× 30

Hugging Face, same model3 models, with then without× 150 to 700

1101001,000

Source: Arcep and PEReN, 2026; Hugging Face, 2025 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

An agent receives a goal rather than a question. It calls the model, which decides on an action: reading a file, running a search, executing code. The result is added to the context, and the model is called again with the whole, which grows longer at each step. The text reread thus grows from call to call. Anthropic states that its agents consume about 4 times more tokens than a conversation, and its multi-agent systems about 15 times more. On software repair tasks, Bai and co-authors count 1,000 times more tokens for an agent than for a code exchange or reasoning about code, mostly input tokens, that is reread context.

The climatologist Zeke Hausfather logged his own use of Claude Code over 8 weeks in 2026: 1,138 instructions typed, more than 14,000 calls to the model and 3.2 billion tokens. 96% of these tokens are rereads of the context, held in cache, and the text he sees is only 0.4% of the total. According to him, this amounts to 12 calls and 2.9 million tokens on average per instruction. He estimates its energy at 150 Wh per instruction, from 60 to 290 Wh depending on the method for converting tokens into energy and the cost assigned to a cached reread (10% of that of a fresh read in his central estimate, from 1 to 25% in his bounds); this is the estimate of a heavy user, not a study. Microsoft estimates that long reasoning queries and agent queries consume more than 10 times the energy of a median query, and the IEA writes that reasoning, agents and video generation can consume hundreds or thousands of times the energy of a simple text generation, without detailing the calculation. Other published estimates, lower, are detailed in our article on the footprint of Claude↗.

From a short response estimated at 0.3 Wh to an agent instruction estimated at 150 Wh, the ratio is about 500, our calculation; Zeke Hausfather himself counts about 600 times a median question. An agent instruction is not a question: it triggers 12 calls to the model on average. The gap comes first from the number of tokens: a longer draft, then a context reread at each call. Over a day, the heavy agent use logged by Zeke Hausfather amounts to about 3 kWh; the analysis of a year of use derives the annual order of magnitude from it, the analysis of tasks by occupation places it next to other uses, and the one on everyday actions quantifies what the choice of mode changes.

For us at Projet Celsius, an energy figure "per query" says little if it does not specify the mode. The short response, around 0.3 Wh, is the best-documented and least costly use; reasoning and agents multiply it by 2 to several hundred times depending on the model, the task and the available estimates, and these are the ones spreading. Providers do not publish the energy of their reasoning modes or their agents: the orders of magnitude above come from open models, external estimates and one user's log. For the user, the lever is the same in all 2 cases: the amount of text the model is made to write and reread.