23.2 Wh
#1
energy per medium-length prompt for DeepSeek V3.2 Exp, the highest of the 2025 models measured

What the chart shows

Jegham et al. (2025) estimated the energy each model uses to answer a medium-length prompt, about 1,000 tokens in and 1,000 out. Among the 15 most energy-intensive models of 2025, DeepSeek V3.2 Exp (23.2 Wh), DeepSeek V3.2 (23.1 Wh), GPT-5 at high reasoning effort (21.9 Wh) and o3-pro (21.8 Wh) used the most.

The same model can vary widely with its settings: GPT-5 used 21.9 Wh at high reasoning effort, 13.1 Wh at medium and 8.4 Wh at low. Claude 4 Opus (5.3 Wh) and GPT-5 mini at medium effort (5.1 Wh) were the lowest of these 15.

Key numbers

MeasureValue
DeepSeek V3.2 Exp23.24 Wh
GPT-5 (high)21.85 Wh
GPT-5 (medium)13.08 Wh
GPT-5 (low)8.35 Wh
Claude 4 Opus5.32 Wh

How to read it

  • These are the 15 highest-energy 2025 models in the study, not all models; many lighter models use less.
  • Estimates from public API latency, hardware assumptions and data centre efficiency factors, not measured power.
  • The study also covers 2023 and 2024 models; download the CSV for all 33.
Source: Jegham et al., 2025, via the AI Index Report 2026 (Stanford HAI), figure 1.4.5: Model energy consumption for medium-length prompts. Chart and text by AI Adoption.
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart