Baseline (standard text)
0,30 Wh
Historical reference to compare everything else
Back to reports
Sustainability
State of AI Energy Consumption(2025-2026)
Dashboard to explore consumption, models, modalities, multipliers and real-world equivalences.
One label, radically different consumption
This panel summarises the key point of the original report: there is no single figure for "AI consumption". Lightweight text remains in the range of tenths of a Wh, while reasoning, agents and video now operate on completely different scales.
Average reasoning
35,4 Wh
Quickly enters 100x orders of magnitude
Video generation
520 Wh
Per short clip in the 2026 commercial range
Code agents
85 Wh
Long sessions with tool calls and massive context
Energy disparity by modality
Logarithmic scale. The jump from text to video is exponential.
API cost vs estimated consumption
Bubble = approximate parametric magnitude. Price does not always reflect physical cost well, but it helps to see trends.
Model directory 2025-2026
Explore the report database: sort by model, category, energy or multiplier and filter by modality.
| Everyday equivalence | Source / confidence | ||||
|---|---|---|---|---|---|
| GPT-4o OpenAI | Text / Reasoning | 0.3 Wh | x1 | 1 sec microwave | Epoch AI |
| Gemini 1.5 Pro (Mediana) Google | Text / Reasoning | 0.24 Wh | x0.8 | 1 sec microwave | Google (Direct) |
| Llama 3.1 8B Meta | Text / Reasoning | 0.03 Wh | x0.1 | LED bulb 1 sec | ML.Energy |
| Llama 3.1 405B Meta | Text / Reasoning | 1.4 Wh | x4.7 | 5 sec microwave | ML.Energy |
| o3-reasoning (Consulta Larga) OpenAI | Text / Reasoning | 39.2 Wh | x130.7 | 30 min watching TV | Jegham et al. |
| DeepSeek-R1-Distill-70B DeepSeek | Text / Reasoning | 4.6 Wh | x15.3 | 15 sec microwave | Extrapolated (x154) |
| Phi-4-reasoning-plus Microsoft | Text / Reasoning | 15.4 Wh | x51.3 | Full phone charge | Direct measurement (x514) |
| GPT-5-main OpenAI | Text / Reasoning | 4.5 Wh | x15 | 15 sec microwave | URI AI Lab |
| GPT-5-thinking OpenAI | Text / Reasoning | 35 Wh | x116.7 | 2.5 phone charges | Shaolei Ren (x5-x10) |
| GPT-5.4 OpenAI | Text / Reasoning | 42 Wh | x140 | 1 hour watching TV | Estimated from API pricing |
| Claude Opus 4.6 Anthropic | Text / Reasoning | 12.5 Wh | x41.7 | 40 sec microwave | Thermal estimate |
| Claude Sonnet 4.6 (Adaptive High) Anthropic | Text / Reasoning | 18 Wh | x60 | 1.2 phone charges | Thinking usage analysis |
| Gemini 2.5 Flash Google | Text / Reasoning | 0.15 Wh | x0.5 | Camera flash | |
| Gemini 3.1 Pro Preview Google | Text / Reasoning | 8.5 Wh | x28.3 | 30 sec microwave | v6 hardware estimate |
| Llama 4 Maverick (MoE) Meta | Text / Reasoning | 0.8 Wh | x2.7 | 3 sec microwave | 17B active-params analogy |
| Grok 4 (Gas Powered) xAI | Text / Reasoning | 22 Wh | x73.3 | High carbon footprint | Southern Env. Law Center |
| DeepSeek-V3.2 DeepSeek | Text / Reasoning | 1.2 Wh | x4 | 4 sec microwave | Sparse attention |
| DeepSeek-R1 (HW Ascend) DeepSeek | Text / Reasoning | 33.6 Wh | x112.0 | 2 phone charges | Jegham et al. |
| GPT-4o Image Auto OpenAI | Image | 9.5 Wh | x31.7 | LED bulb 1 hour | Scope3 |
| SDXL H100 Stability | Image | 1.64 Wh | x5.5 | 10 min LED bulb | HuggingFace AI Energy |
| Midjourney v7 Midjourney | Image | 12 Wh | x40 | 1 phone charge | Community estimate |
| AudioLDM Various | Audio | 0.25 Wh | x0.8 | 1 sec microwave | Passoni et al. |
| Suno v5.5 (Canción) Suno | Audio | 25 Wh | x83.3 | 2 phone charges | GPU-time estimate |
| Sora 2 (Clip 10s) OpenAI | Video | 513 Wh | x1,710 | 1 washing machine cycle | Post-launch reports |
| CogVideoX1.5-5B THUDM | Video | 944 Wh | x3,146.7 | Half dishwasher cycle | MIT Tech Review |
| Veo 3 Standard (10s) Google | Video | 180 Wh | x600 | 1 km electric car | Cost-based estimate |
| Kling 3.0 (15s) Kuaishou | Video | 145 Wh | x483.3 | 1 km electric car | Infrastructure estimate |
| Claude Code (Sesión) Anthropic | Agents / Code | 41 Wh | x136.7 | 30 min watching TV | Simon P. Couch |
| Devin 2.0 (Tarea) Cognition | Agents / Code | 120 Wh | x400 | 8 phone charges | Efficiency +83% vs v1 |
| Aider (Sesión) Various | Agents / Code | 9.8 Wh | x32.7 | 10 min watching TV | Token efficiency (105k) |
| OpenAI o3 Deep Research OpenAI | Agents / Code | 450 Wh | x1,500 | Almost 1 washing machine cycle | Artificial Analysis |
| Perplexity Deep Research Perplexity | Agents / Code | 85 Wh | x283.3 | 1 hour watching Netflix | Extrapolation from 308k tokens |
Baseline: 0.30 Wh as a standard text query.
Real impact calculator
Convert the abstract operation into a tangible equivalence: phone charges, hours of TV, kilometres in an electric car or washing machine cycles.
Estimated total consumption
9,440 Wh
Smartphone charges
674.29 Hours watching Netflix
122.6 Km in electric car
62.93 Washing machine cycles
18.88 Consumption is too low to show these macro equivalences.
AISHA reading: generative video, deep research and extended agents are not simple “features”. They are architecture and material expenditure decisions. The right approach is not to ban them, but to measure when it is worth activating them.
Sources
- AI's Power Demand: Calculating ChatGPT's electricity consumption
- Your AI Carbon Footprint: What Every Query Really Costs
- Electricity use of AI coding agents — Simon P. Couch
- Google Gemini 3.1 Pro Review and Analysis
- OpenAI Sora Shutdown: $15M/Day Costs, $2.1M Revenue
- Meta Llama 4 Models: Features, Benchmarks
- DeepSeek 3.2 and efficient open AI
- Inside Memphis' Battle Against Elon Musk's xAI Data Center