What is the carbon footprint of Llama and Meta AI, and how much energy does a response use?
For sustainability teams and engineers who need a number for the Llama models their company runs or calls. Sources checked on 20 November 2026. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.
Page updated
Part of the SOMA AI Emission Factors cluster.
How much energy does a Llama response use?
About 0.032 Wh (114 J) for a response from Llama 3.1 8B and about 1.86 Wh (6,706 J) for one from Llama 3.1 405B, on the estimates in MIT Technology Review's article of 20 May 2025. The larger model uses about 59 times the energy of the smaller one. These are estimated totals for a response, with the article's own boundary, not figures published by Meta. We found no per-query figure from Meta for Llama or for the Meta AI assistant.
| Figure | Source | Wh per response |
|---|---|---|
| Llama 3.1 8B, 114 J estimated total | MIT Technology Review, 20 May 2025 | 0.032 |
| Llama 3.1 405B, 6,706 J estimated total | MIT Technology Review, 20 May 2025 | 1.86 |
| Llama 3.1 405B FP8, mean request at one heavily batched H100 setting | ML.ENERGY Leaderboard, v3.0, data as of 16 February 2026 | 0.21 |
The ML.ENERGY row is GPU energy only, at one heavily batched setting, so compare it with GPU figures and not with the doubled totals.
Can you measure Llama yourself?
Llama is open-weight, so you can run it on your own hardware and meter it. That is the case where the user can measure, and not only estimate, an AI model's energy. A measurement of your own server beats a class estimate, provided you state what was metered, which is the GPU, the whole server or the data centre. The steps are in how to measure the environmental impact of AI.
What if you call Llama through a cloud API?
Then you cannot meter it, and SOMA AI (somaai.earth) estimates it, from the open SOMA AI Emission Factors (SAEF). SAEF assigns each Llama model a class from its parameter count and applies the grid of the serving region. The default region, us-east-1, is an assumption, so use the region in the usage export when there is one.
| Model | SAEF class | Wh | us-east-1 (US East - N. Virginia) | eu-north-1 (EU North - Sweden) | eu-west-1 (EU West - Ireland) |
|---|---|---|---|---|---|
| Llama 3.1 8B | Class A | 0.018 | 0.0065 g | 0.0017 g | 0.0061 g |
| Llama 3.1 70B | Class B | 0.12 | 0.043 g | 0.01 g | 0.04 g |
| Llama 3.1 405B | Class C | 0.37 | 0.13 g | 0.037 g | 0.13 g |
The 405B model is 21 times the 8B model in the same region, and the same 405B exchange is 3.6 times lower in EU North - Sweden than in US East - N. Virginia. The estimates are for an exchange of 928 input and 282 output tokens, a different unit from the per-response figures above.
How does Llama compare with ChatGPT, Claude and Gemini?
SAEF gives models of the same class and the same region the same number, so it cannot show whether one brand is lower than another. The comparison below is for illustration at one assumed region.
| SAEF class | Llama | Claude | OpenAI | Google Gemini |
|---|---|---|---|---|
| Class A | llama-3.1-8b, 0.0065 g | claude-haiku-4, 0.01 g | gpt-5-mini, 0.01 g | gemini-2.5-flash, 0.01 g |
| Class B | llama-3.1-70b, 0.043 g | claude-sonnet-4, 0.036 g | gpt-4o, 0.036 g | gemini-2.5-pro, 0.036 g |
| Class C | llama-3.1-405b, 0.13 g | claude-opus-4, 0.13 g | gpt-5, 0.13 g | No Gemini model in this class |
More models are on the model comparison.
What should a company report for Llama?
If you self-host, report the metered energy, what the meter covered, and the grid you ran on. If you call Llama through a cloud API, take token counts per model from the usage export and report the class, the region, the factor version and the range, and say that Meta has published no per-query figure that we could find. The method is in the Scope 3 guide and on the methodology page, and the free calculator gives a first estimate without an account.
Short answers
- How much energy does a Llama response use?
- MIT Technology Review's estimate is 114 J, about 0.032 Wh, for a response from Llama 3.1 8B and 6,706 J, about 1.86 Wh, for one from Llama 3.1 405B. That is a ratio of about 59 times between the two sizes. We found no per-query figure published by Meta.
- Does Meta AI have a published carbon footprint per response?
- We found no published per-query figure from Meta, so any figure for Meta AI is an estimate. SAEF can estimate the open Llama models by size class and serving region, but it does not know where or on what hardware Meta runs its own assistant.
- Can I measure Llama's energy myself?
- Yes, if you run it yourself. Open-weight models can be self-hosted, so you can meter the GPU or the server directly and report a measurement. If you call Llama through a cloud API, you can only estimate it by size class and region.
- Is Llama lower or higher than ChatGPT, Claude and Gemini?
- SAEF cannot tell providers apart inside a size class. It assigns Llama a class from its parameter count and every other model a class from its provider's tier, so same class and same region give the same number. A measured figure for a Llama model on your own hardware is a different kind of evidence.
Keep reading
- How do you measure the environmental impact of AI? Meter what you run, estimate what you call
- How much energy does generating an AI image or video use? Image and video against Llama text responses
- Which AI model has the lowest carbon footprint? GPT, Claude, Gemini, Llama, DeepSeek and Mistral compared. Twelve models side by side
- What is the carbon footprint of Claude, and how does it compare with ChatGPT? Haiku, Sonnet and Opus by region
