Reasoning model or standard model: how much does the bill go up?
On the same prompt, a reasoning model uses 2 to 6.9 times more electricity than a standard model from the same vendor. 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 DeepSeek-V3 and DeepSeek-R1, on the same host. On a long prompt, R1 uses 29.08 Wh against 13.16 Wh for V3 at DeepSeek. The "65 times more" in the same study compares something else: a reasoning model against a small 8-billion-parameter model.
Reasoning multiplies consumption by 2 to 6.9 depending on prompt size, not by 65. The gap is widest on a short prompt (6.9 times at DeepSeek) and narrowest on a long one (2.2 times). The host matters more than the mode: on a long prompt, the same R1 uses 3.9 times more electricity at DeepSeek than on Azure.
Figures updated on October 7, 2026
| Host | Prompt | DeepSeek-V3 | DeepSeek-R1 | R1 ÷ V3 |
|---|---|---|---|---|
| DeepSeek | Short (100 tokens in, 300 out) | 2.78 Wh | 19.25 Wh | 6.9 × |
| DeepSeek | Medium (1,000 and 1,000) | 8.86 Wh | 24.60 Wh | 2.8 × |
| DeepSeek | Long (10,000 and 1,500) | 13.16 Wh | 29.08 Wh | 2.2 × |
| Microsoft Azure | Short (100 tokens in, 300 out) | 0.74 Wh | 2.35 Wh | 3.2 × |
| Microsoft Azure | Medium (1,000 and 1,000) | 2.17 Wh | 4.33 Wh | 2.0 × |
| Microsoft Azure | Long (10,000 and 1,500) | 3.70 Wh | 7.41 Wh | 2.0 × |
Source: Table 4 of Jegham et al., arXiv 2505.09598 (version 6, 24 November 2025). The energy figures are the table's. The "R1 ÷ V3" 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 does reasoning add to consumption, on the same prompt?
Between 2 and 6.9 times, depending on prompt size and host. At DeepSeek, on a short prompt (100 tokens in, 300 out), R1 uses 19.25 Wh against 2.78 Wh for V3: 6.9 times. On a medium prompt, 2.8 times. On a long prompt, 2.2 times. On Azure, the ratios are 3.2, 2.0 and 2.0. The DeepSeek pair is the only one in the table that isolates reasoning: two models from one vendor, in the same place, on the same prompts. The study's other reasoning models (o1, o3, o4-mini) have no standard twin. Comparing them with some other model would blend the effect of reasoning with the effect of size.
Why is it not the study's "65 times"?
Because that 65 compares two different things. The study states that DeepSeek-R1 at DeepSeek uses 29 Wh on a long prompt, about 65 times the most efficient model tested, LLaMA-3.1-8B (0.443 Wh). That model has 8 billion parameters, so the gap adds up model size, host and reasoning. Reasoning alone weighs 2.2 times on that same prompt. The first version of the study (14 May 2025) reported over 33 Wh and more than 70 times GPT-4.1 nano. Version 6 (24 November 2025) corrected those values.
Why does the gap shrink as the prompt gets longer?
The study does not explain it. The figures, though, are clear. At DeepSeek, V3 goes from 2.78 to 13.16 Wh as the prompt gets longer, 4.7 times more. R1 goes from 19.25 to 29.08 Wh, 1.5 times more. The ratio between the two therefore falls from 6.9 to 2.2. One possible reading: a long prompt weighs heavily on both models, and R1's extra thinking changes the proportion less. That is a hypothesis, not a result of the study. What to keep is the direction: the reasoning surcharge is highest on short requests.
Does the host matter more than the reasoning mode?
On a long prompt, yes. The same R1 uses 29.08 Wh at DeepSeek and 7.41 Wh on Microsoft Azure: 3.9 times more. The same V3, 13.16 Wh against 3.70 Wh: 3.6 times more. Reasoning adds 2.2 times at DeepSeek and 2.0 times on Azure. The study's authors write that hardware and data center efficiency, not model design alone, drive real-world energy use. Choosing where a model runs therefore changes the bill more than choosing its mode.
Sources
Frequently asked
How much more does a reasoning model use than a standard model?+
Between 2 and 6.9 times more on the same prompt, according to Table 4 of the study "How Hungry is AI?" (version 6, 24 November 2025) for DeepSeek-R1 and DeepSeek-V3. On a long prompt at DeepSeek: 29.08 Wh against 13.16 Wh.
Why does this site use 20 Wh for reasoning against 0.3 Wh for a text prompt elsewhere?+
Because that ratio, about 67 times, compares a reasoning request with an average text prompt, not with a standard model on the same prompt. It is used to compute the tipping point for a whole service, on the page "The reasoning tipping point". The two figures do not answer the same question. On a long prompt, the study's table puts o3 at 12.22 Wh, R1 at 29.08 Wh at DeepSeek and 7.41 Wh on Azure: the site's 20 Wh is a choice in the middle of that spread.
Do these figures apply to ChatGPT, Claude or Gemini?+
No. The pair tested is DeepSeek, and the study estimates consumption without access to the servers, from public API performance and assumed hardware. These are orders of magnitude for specific models at a specific date, not a rule for every assistant.
Are these ratios in the study?+
No. Table 4 gives mean energy in Wh per model and per prompt size, with a standard deviation. The ratios on this page are a division done by this site on those means. The standard deviation is large for R1 at DeepSeek (± 9.4 Wh on the short prompt): a ratio of 6.9 is an order of magnitude, not an exact value.
A reasoning model costs 2 to 7 times more than a standard model on the same prompt. Whether the question deserves that spend is a call about use, not arithmetic. The studio GHIS helps frame that call before an assistant is deployed.