26 entries
Model footprints
Only three providers have published a per-prompt figure, and their scopes differ: Google counts on-site cooling water, Mistral counts the water behind the electricity too. Every other model carries an EcoGPT estimate with its working attached — hover the labels to see which is which.
How to read the labels:Company figureCompany figure. Published by the company itself. We take it as given and record the scope next to it.Measured independentlyMeasured independently. Measured or estimated by researchers or analysts with no stake in the result.ecogpt.co estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement.DerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor.Not disclosedNot disclosed. The company publishes nothing on this and no defensible estimate exists.— hover any label.
| Model | Provider | Weights | Energy / prompt | Water / prompt | CO₂e / prompt | Training CO₂e | Sources |
|---|---|---|---|---|---|---|---|
Apple Foundation Models (on-device) ≈3B on-device; server model on Apple silicon · 2025 |  Apple | Closed | 0.020 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.004 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.008 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 1,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Luccioni, Jernite, Strubell (arXiv 2311.16863)Apple |
Open text-generation models (median, 2023 study) Sub-10B open models · 2023 | Independent research | Open | 0.042 WhIndependentMeasured independently. Measured or estimated by researchers or analysts with no stake in the result. Median of 0.042 kWh per 1,000 inferences; image generation median 1.35 kWh per 1,000. | 0.076 mLDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.016 gDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. | 50.0 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Luccioni, Jernite, Strubell (arXiv 2311.16863)Lawrence Berkeley National Laboratory |
Granite 4.0 (Micro / Small) 3B–32B hybrid Mamba/transformer · 2025 |  IBM | Open | 0.050 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.050 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.019 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 300 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Luccioni, Jernite, Strubell (arXiv 2311.16863)IBMIBM Community |
Llama 4 Scout 109B total / 17B active (MoE) · 2025 |  Meta | Open | 0.100 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.020 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.039 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 1,354 tDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. Market-based reported as 0. | Meta (model card)Epoch AIMeta |
Llama 4 Maverick 400B total / 17B active (MoE) · 2025 | Meta | Open | 0.150 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.030 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.058 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 645 tDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | Meta (model card)Epoch AIMeta |
DeepSeek-V3 671B total / 37B active (MoE) · 2024 |  DeepSeek | Open | 0.200 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.116 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 1,455 tDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. DeepSeek disclosed the GPU-hours (≈$5.6M at $2/GPU-hour) but not emissions; H800 at 0.7 kW plus cooling and host. | DeepSeek (arXiv 2412.19437)Epoch AIKoomey Analytics |
Qwen 3 0.6B–235B (dense + MoE) · 2025 |  Alibaba | Open | 0.200 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.116 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 3,400 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Hugging Face, Salesforce, Cohere, Carnegie MellonEpoch AIKoomey AnalyticsAlibaba Group |
Doubao 1.5 Pro MoE (≈20B active, undisclosed total) · 2025 |  ByteDance | Closed | 0.200 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.116 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 3,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Analytics InsightEpoch AIKoomey AnalyticsDeepSeek (arXiv 2412.19437) |
Kimi K2 Thinking 1T total / 32B active (MoE) · 2025 Training cost figure is unofficial; the CEO says it excludes research and experiments. |  Moonshot | Open | 0.200 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.116 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 1,500 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | ImplicatorEpoch AIKoomey AnalyticsYicai Global |
Gemini (Apps, median text prompt) Frontier (undisclosed) · 2025 Fleet-wide measurement, May 2025. Energy per prompt fell 33× in twelve months. |  Google | Closed | 0.240 WhDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. Comprehensive method incl. idle capacity; 0.10 Wh active-chip only. | 0.260 mLDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. On-site cooling only. | 0.030 gDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | 9,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Google (arXiv 2508.15734)Google |
Llama 3.1 family (8B / 70B / 405B) 8B–405B dense · 2024 | Meta | Open | 0.250 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.050 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.098 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 11,390 tDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | Meta (model card)Luccioni, Jernite, Strubell (arXiv 2311.16863)Epoch AIMeta |
ERNIE 4.5 424B total / 47B active (MoE) · 2025 |  Baidu | Open | 0.250 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.145 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 4,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Reuters via Yahoo FinanceEpoch AIUS Securities and Exchange Commission |
Hunyuan T1 / Turbo S 389B total / 52B active (MoE) · 2025 |  Tencent | Closed | 0.250 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.200 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.145 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 4,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | TencentEpoch AI |
Amazon Nova Pro Mid-size (undisclosed) · 2024 |  Amazon | Closed | 0.250 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.037 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.098 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 3,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | AmazonLuccioni, Jernite, Strubell (arXiv 2311.16863)Epoch AIAmazon |
GLM-4.6 355B total / 32B active (MoE) · 2025 |  Zhipu | Open | 0.250 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.250 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.145 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 5,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | eWeekEpoch AIKoomey AnalyticsDolphin Research (Longbridge)KuCoin News |
GPT-4o (typical query, independent estimate) Frontier (undisclosed) · 2024 |  OpenAI | Closed | 0.300 WhIndependentMeasured independently. Measured or estimated by researchers or analysts with no stake in the result. ≈2.5 Wh for a 10k-token input; ≈40 Wh at 100k tokens. | 0.090 mLDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.117 gDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. | 10,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Epoch AIMicrosoft |
Claude (Opus / Sonnet / Haiku) Frontier (undisclosed) · 2025 |  Anthropic | Closed | 0.300 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.045 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.117 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 10,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Epoch AIAmazonAmazon |
Grok 3 / 4 Frontier (undisclosed) · 2025 |  xAI | Closed | 0.300 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.180 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.150 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 150,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Epoch AISemiAnalysisFOX13 MemphisInside Climate News |
Nemotron (Ultra / Super) 49B–253B (Llama-derived) · 2025 |  NVIDIA | Open | 0.300 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.090 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.117 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 5,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Epoch AINVIDIAMicrosoft |
Agentforce (typical agent turn) Partner frontier models + xGen small models · 2024 |  Salesforce | Closed | 0.300 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.045 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.117 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | — Not applicable — runs on partner models; Salesforce's own xGen models are small. | Techstrong.aiEpoch AILuccioni, Jernite, Strubell (arXiv 2311.16863)Amazon |
ChatGPT (average query, company figure) Frontier (undisclosed) · 2025 | OpenAI | Closed | 0.340 WhDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. Blog post, no methodology. | 0.320 mLDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. 0.000085 US gallons; scope unstated. | 0.133 gDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. | 15,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Sam Altman (blog)Epoch AI |
Command A 111B dense · 2025 Open weights for research; Cohere co-created the AI Energy Score. |  Cohere | Open | 0.400 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.120 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.156 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 2,000 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | The LogicEpoch AIMicrosoft |
Mistral Large 2 (Le Chat, 400-token response) 123B dense · 2024 Peer-reviewed LCA with Carbone 4 and ADEME. |  Mistral | Open | 0.600 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 45.0 mLDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. Full scope incl. electricity-generation water. | 1.140 gDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | 20,400 tDisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | Mistral AIEpoch AI |
Perplexity Sonar (search answer) Llama 3.3 70B fine-tune + retrieval · 2025 |  Perplexity | Closed | 0.800 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 0.120 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.312 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 500 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | CerebrasEpoch AIAmazon |
Stable Diffusion XL (per image) 3.5B diffusion · 2023 | Independent research | Open | 1.640 WhIndependentMeasured independently. Measured or estimated by researchers or analysts with no stake in the result. AI Energy Score; 11.49 Wh with the model's default (larger) output settings per Luccioni et al. | 2.952 mLDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. On-site cooling only, so comparable with Google and OpenAI's figures. | 0.640 gDerivedDerived. Arithmetic on a disclosed or independent input using a published conversion factor. | 25.0 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Hugging Face, Salesforce, Cohere, Carnegie MellonLuccioni, Jernite, Strubell (arXiv 2311.16863)Lawrence Berkeley National Laboratory |
Midjourney V7 (per image) Diffusion (undisclosed) · 2025 |  Midjourney | Closed | 2.900 WhEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; no provider figure. | 3.335 mLEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. On-site cooling only, so comparable with Google and OpenAI's figures. | 1.131 gEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 500 tEstimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. ecogpt.co estimate; the provider has not disclosed training emissions. | Hugging Face, Salesforce, Cohere, Carnegie Mellonheise onlineGoogle (arXiv 2508.15734)Luccioni, Jernite, Strubell (arXiv 2311.16863) |