The AI Footprint

Which AI models are most energy efficient?

What it actually costs the planet to use one AI model instead of another. Every figure here is calculated the same way, from public data, and checked against laboratory measurements.
First model
Grok 4.7
against
Second model
GPT-5.4 Pro
Compare
Answer, as of
o4-mini
OpenAI · capability 145.6
0.258 Wh per answer
0.12 g CO₂·under a metre of driving
44 models sit within 1.5× of the lowest energy figure here, closer than our accuracy can separate, so we will not rank them on energy. Of those, this one scores highest on Epoch's Capabilities Index, a composite over 58 benchmarks covering coding, agents, long context and writing, among models people have actually voted on. It carries 44,855 head-to-head human votes on LMArena. Grok 4.7 scores marginally higher on the benchmarks (154.37 against 145.6) but has never faced a human vote, so we do not put it first. Calculated from an estimated parameter count, the maker hasn't published one.
Best of the top-tier models
Grok 4.7
xAI
capability 154.4 · this band is 152+
0.285 Wh
0.13 g CO₂·under a metre of driving
62× less energy than the typical model of this standard. Calculated from a published parameter count.
Best of the strong all-rounders
o4-mini
OpenAI
capability 145.6 · this band is 143+
0.258 Wh
0.12 g CO₂·under a metre of driving
29× less energy than the typical model of this standard. Calculated from an estimated parameter count, the maker hasn't published one.
Best of the solid everyday models
o3-mini
OpenAI
capability 140.3 · this band is 130+
0.248 Wh
0.11 g CO₂·under a metre of driving
The lightest model at this standard. Calculated from an estimated parameter count, the maker hasn't published one.
Released 2026-10-15
Meta AI
No figure yet
Independent testers at Epoch AI have published 5 early results, but not the overall score our energy estimate is built from. Any figure now would be a guess, so we show none. It fills in by itself when the score does.
Released 2026-10-12
Meta AI
No figure yet
Independent testers at Epoch AI have published 6 early results, but not the overall score our energy estimate is built from. Any figure now would be a guess, so we show none. It fills in by itself when the score does.
The early verdict from people. In blind voting on LMArena, 1 of 416 models are rated clearly above it. That rests on 3,247 votes so far, so it will move as more come in.
Same capability. 2790× the energy.

Among the 40 models scoring Top tier (152+) across 58 benchmarks, the heaviest uses about 2790× the energy of the lightest, for work of the same standard. Choosing well costs you almost nothing in capability.

Lightest in this band
Grok 4.7
xAI
0.285 Wh
0.13 g CO₂·under a metre of driving
capability 154.4
2790×
more energy
Heaviest in this band
GPT-5.4 Pro
OpenAI
796 Wh
365 g CO₂·≈ 1.5 km (0.9 miles) of driving
capability 159.0
Compare any two models
GPT-5.4 Pro
OpenAI · capability 159.0
796 Wh
365 g CO₂·≈ 1.5 km (0.9 miles) of driving
About 500× the typical model here. Calculated from an estimated parameter count, the maker hasn't published one.
GPT-5.5 Pro
OpenAI · capability 162.0
730 Wh
334 g CO₂·≈ 1.3 km (0.8 miles) of driving
About 450× the typical model here. Calculated from an estimated parameter count, the maker hasn't published one.
GPT-5.2 Pro
OpenAI · capability 156.0
670 Wh
307 g CO₂·≈ 1.2 km (0.8 miles) of driving
About 420× the typical model here. Calculated from an estimated parameter count, the maker hasn't published one.
Most capable
Claude Opus 5.5
Anthropic · capability 166.8 · 17.0 Wh per answer. Ranks 236 of 277 on energy, using about 69× the energy of the most efficient model tracked.
Energy cannot be worked out without knowing how big a model is. 8 of the 285 models here publish no size at any source, including several of the most capable ever released. This is not a gap in our data so much as a gap in theirs.
Technology Innovation Institute
0% · 4 models
Hugging Face +3
0% · 3 models
Baichuan
0% · 3 models
DeepSeek +1
0% · 3 models
xAI
27% · 11 models
Moonshot
33% · 6 models
MiniMax
33% · 3 models
Meta AI
50% · 24 models
Microsoft
50% · 6 models
Alibaba
53% · 45 models
Anthropic
54% · 26 models
01.AI
67% · 3 models
DeepSeek
69% · 16 models
Google DeepMind
75% · 28 models
Z.ai (Zhipu AI)
83% · 6 models
OpenAI
87% · 39 models
Mistral AI
94% · 18 models
enough published to assesspartly published, wide bracket onlynothing published, cannot assess
ModelMakerCapabilityEnergy / answer ↑ConfidenceThinkingReleased
o3-miniOpenAI140.30.248 WhDay-one estimateHidden2025-01-31
Gemini 2.5 Flash-Lite (Jun 2025)Google DeepMind133.90.249 WhDay-one estimateHidden2025-06-17
o4-miniOpenAI145.60.258 WhDay-one estimateHidden2025-04-16
GPT-5 nanoOpenAI139.30.259 WhDay-one estimateHidden2025-08-07
GPT-4o miniOpenAI126.60.264 WhDay-one estimateNone2024-07-18
stablelm-tuned-alpha-7b56.00.280 WhInferredNot stated2023-04-19
DeepSeek Coder 1.3BDeepSeek +163.90.280 WhInferredShown2023-11-02
Qwen2.5-Coder-0.5B88.30.280 WhInferredNot stated2024-09-18
StarCoder 2 3BHugging Face +388.90.281 WhInferredShown2024-02-22
Gemma 3 1BGoogle DeepMind-0.282 WhCalculatedShown2025-03-12
Cerebras-GPT-13BCerebras Systems83.40.282 WhInferredShown2023-03-20
DeepSeek Coder 6.7BDeepSeek +189.80.282 WhInferredShown2023-11-02
INTELLECT-1Prime Intellect +2101.00.283 WhInferredNot stated2024-11-29
CodeQwen1.5-7B95.00.283 WhInferredNot stated2024-04-15
Llama 3.2 1BMeta AI102.00.283 WhCalculatedShown2024-09-24
StarCoder 2 7BHugging Face +393.70.283 WhInferredShown2024-02-20
Phi-1.5Microsoft91.70.283 WhCalculatedShown2023-09-11
Qwen-1_8B93.10.284 WhInferredNot stated2023-11-30
Qwen3-1.7BAlibaba-0.284 WhCalculatedShown2025-04-29
DeepSeek-R1-Distill-Qwen-1.5BDeepSeek-0.284 WhCalculatedShown2025-01-20
Grok 4.7xAI154.40.285 WhCalculatedNot stated2026-09-21
Qwen2.5-Coder (1.5B)Alibaba103.10.286 WhInferredShown2024-09-18
Baichuan1-7BBaichuan90.60.286 WhInferredShown2023-06-01
Gemma 2BGoogle DeepMind94.30.286 WhCalculatedShown2024-02-21
Dolly 2.0-12bDatabricks89.80.286 WhInferredShown2023-04-11
RedPajama-INCITE-7B-Base90.60.287 WhInferredNot stated2023-05-04
open_llama_7b91.80.287 WhInferredNot stated2023-06-07
Qwen3-30B-A3BAlibaba136.20.287 WhCalculatedShown2025-04-29
DeepSeek Coder 33BDeepSeek +196.50.288 WhInferredShown2023-11-02
Ministral 3BMistral AI118.10.288 WhCalculatedNot stated2024-10-16
XGen-7BSalesforce93.50.288 WhInferredShown2023-06-27
gpt-oss-20bOpenAI137.80.289 WhCalculatedShown2025-08-05
Baichuan 2-7BBaichuan96.50.289 WhInferredShown2023-09-20
Gemma 3 4BGoogle DeepMind116.00.290 WhCalculatedShown2025-03-12
Qwen3-4BAlibaba-0.290 WhCalculatedShown2025-04-29
Qwen3.5-4BAlibaba-0.290 WhCalculatedShown2026-02-24
Qwen2.5-Coder-3B107.80.292 WhInferredNot stated2024-09-18
MPT-7BMosaicML94.70.293 WhInferredShown2023-05-05
vicuna-13b-v1.194.80.294 WhInferredNot stated2023-04-12
Falcon-7BTechnology Innovation Institute95.20.295 WhInferredShown2023-04-24
Llama 2-7BMeta AI99.20.298 WhCalculatedShown2023-07-18
Eurus-2-7B-PRIMETsinghua University +5-0.298 WhCalculatedShown2024-12-31
Mistral 7BMistral AI-0.298 WhCalculatedShown2024-05-27
Mistral 7B v0.1Mistral AI112.30.298 WhCalculatedShown2023-09-27
Mistral 7B v0.3Mistral AI109.10.298 WhCalculatedShown2024-05-27
Qwen2.5-7BAlibaba118.50.299 WhCalculatedShown2024-09-19
StarCoder 2 15BHugging Face +3105.10.299 WhInferredShown2024-02-20
Llama 3.1-8BMeta AI116.60.300 WhCalculatedShown2024-07-23
Llama 3-8BMeta AI116.50.300 WhCalculatedShown2024-04-18
Gemini 1.5 Flash 8BGoogle DeepMind-0.300 WhCalculatedNone2024-10-03
Ministral 8BMistral AI-0.300 WhCalculatedShown2024-10-16
Qwen3-8BAlibaba136.20.301 WhCalculatedShown2025-04-28
chatglm2-6b99.00.301 WhInferredNot stated2023-06-24
Gemma 7BGoogle DeepMind112.00.302 WhCalculatedShown2024-02-21
DeepSeek-Coder-V2-Lite-Base109.00.302 WhInferredNot stated2024-06-13
Qwen3.5-9BAlibaba139.40.303 WhCalculatedShown2026-02-24
Gemma 2 9BGoogle DeepMind119.80.303 WhCalculatedShown2024-06-24
LLaMA-7BMeta AI96.80.304 WhInferredShown2023-02-24
Falcon 2 11BTechnology Innovation Institute109.60.307 WhInferredShown2024-05-09
Yi 6B01.AI104.90.307 WhInferredShown2023-11-22
MPT-30BMosaicML100.80.307 WhInferredShown2023-06-22
Yi-9B107.70.308 WhInferredNot stated2024-03-01
Nemotron-4 15BNvidia107.80.309 WhInferredNot stated2024-02-26
Baichuan2-13BBaichuan103.30.310 WhInferredShown2023-09-06
Gemma 3 12BGoogle DeepMind123.50.310 WhCalculatedShown2025-03-12
Mistral NeMoMistral AI118.70.310 WhCalculatedShown2024-07-18
Llama 2-13BMeta AI106.20.313 WhCalculatedShown2023-07-18
Phi-4Microsoft Research130.40.316 WhCalculatedShown2024-12-12
phi-3-medium 14BMicrosoft121.20.316 WhCalculatedShown2024-04-23
internlm-7b103.00.317 WhInferredNot stated2023-07-05
Qwen3-14BAlibaba138.20.318 WhCalculatedShown2025-04-29
DeepSeek-R1-Distill-Qwen-14BDeepSeek135.40.318 WhCalculatedShown2025-01-20
Phi-2Microsoft108.00.319 WhInferredShown2023-12-12
Gemma 4 26B A4BGoogle DeepMind141.80.339 WhDay-one estimateShown2026-04-02
Magistral Small 1.0Mistral AI133.20.341 WhCalculatedShown2025-06-10
Mistral Small 3.2Mistral AI131.80.341 WhCalculatedShown2025-06-20
Magistral Small 1.2Mistral AI131.40.341 WhCalculatedShown2025-09-18
Mistral Small 3.1Mistral AI127.50.341 WhCalculatedShown2025-03-17
Mistral Small 3Mistral AI127.10.341 WhCalculatedShown2025-01-30
Gemma 3 27BGoogle DeepMind130.10.349 WhCalculatedShown2025-03-12
Gemma 2 27BGoogle DeepMind122.10.349 WhCalculatedShown2024-06-24
Qwen3.6 27BAlibaba146.50.351 WhCalculatedShown2026-04-22
Gemma 4 31B ITGoogle DeepMind142.90.359 WhDay-one estimateShown2026-04-02
DeepSeek-R1-Distill-Qwen-32BDeepSeek137.40.362 WhCalculatedShown2025-01-20
Qwen1.5-32BAlibaba-0.362 WhCalculatedShown2024-04-03
QwQ-32BAlibaba137.60.363 WhCalculatedShown2025-03-05
Qwen2.5-32BAlibaba128.50.363 WhCalculatedShown2024-09-17
Qwen3-32BAlibaba138.50.364 WhCalculatedShown2025-04-29
Qwen2.5-Coder-14B116.40.448 WhInferredNot stated2024-09-18
LLaMA-13BMeta AI100.70.459 WhInferredShown2023-02-24
Qwen-7BAlibaba106.90.461 WhInferredShown2023-09-28
Llama 2-34BMeta AI105.30.466 WhInferredNot stated2023-07-18
Qwen2.5-Coder-32BAlibaba119.50.469 WhInferredShown2024-09-18
Command R+Cohere +1119.40.473 WhInferredShown2024-08-30
Falcon-40BTechnology Innovation Institute104.50.475 WhInferredShown2023-05-25
GPT-6 LunaOpenAI156.30.492 WhDay-one estimateHidden2026-09-22
PaLM 2-S106.30.495 WhInferredNot stated2023-05-17
Claude Haiku 5.5Anthropic159.00.515 WhDay-one estimateHidden2026-10-07
Claude Haiku 4.5Anthropic142.40.550 WhDay-one estimateHidden2025-10-15
GPT-4.1 nanoOpenAI129.60.551 WhDay-one estimateHidden2025-04-14
Qwen 3.6 35B-A3BAlibaba144.10.575 WhCalculatedShown2026-04-14
Qwen 3.5 Flash (hosted 35B-A3B)Alibaba143.80.575 WhCalculatedNot stated2026-02-25
Qwen3.5-35B-A3BAlibaba142.50.575 WhCalculatedShown2026-02-24
Mixtral 8x7BMistral AI118.50.583 WhDay-one estimateShown2023-12-11
GPT-5.4 NanoOpenAI146.10.603 WhDay-one estimateHidden2026-03-17
Amazon Nova ProAmazon123.80.687 WhInferredNot stated2024-12-03
phi-3-mini 3.8BMicrosoft117.40.701 WhInferredShown2024-04-23
Gemini 1.5 FlashGoogle DeepMind122.60.704 WhInferredNone2024-09-24
internlm-20b112.20.733 WhInferredNot stated2023-09-18
Yi-34B01.AI117.40.734 WhCalculatedShown2023-11-22
Yi-1.5-34B01.AI-0.734 WhCalculatedShown2024-05-13
PaLM 2-M108.40.739 WhInferredNot stated2023-05-17
Falcon-180BTechnology Innovation Institute112.20.742 WhInferredShown2023-09-06
Claude 3 HaikuAnthropic118.40.747 WhInferredNone2024-03-07
Qwen-14BAlibaba113.10.748 WhInferredShown2023-09-24
LLaMA-33BMeta AI107.50.778 WhInferredShown2023-02-24
Qwen3.7 FlashAlibaba144.60.998 WhInferredNot stated2026-07-27
GPT-5.5 InstantOpenAI142.41.01 WhInferredHidden2026-05-05
Qwen 3.6 FlashAlibaba143.11.03 WhInferredNot stated2026-04-27
Claude 3.5 HaikuAnthropic127.21.10 WhInferredNone2024-10-22
Gemini 1.0 ProGoogle DeepMind117.11.10 WhInferredNone2023-12-13
phi-3-small 7.4BMicrosoft121.81.12 WhInferredShown2024-04-23
Claude 3 SonnetAnthropic120.71.14 WhInferredNone2024-02-29
Grok-2xAI130.51.16 WhInferredNot stated2024-12-12
gpt-oss-120bOpenAI140.01.17 WhCalculatedShown2025-08-05
LLaMA-65BMeta AI110.21.21 WhInferredShown2023-02-24
DeepSeek-V2 (MoE-236B, May 2024)DeepSeek124.81.22 WhInferredShown2024-05-07
Claude 2.1Anthropic119.31.22 WhInferredNone2023-11-21
Qwen3.5-122B-A10BAlibaba-1.22 WhCalculatedShown2026-02-24
Llama 4 ScoutMeta AI129.71.29 WhCalculatedShown2025-04-05
GPT-5 miniOpenAI145.41.31 WhDay-one estimateHidden2025-08-07
PaLM 2-L115.11.31 WhInferredNot stated2023-05-17
Stable Beluga 2Stability AI117.21.31 WhInferredShown2023-07-20
Gemini 3.5 Flash-LiteGoogle DeepMind145.21.46 WhDay-one estimateHidden2026-07-21
DBRXDatabricks-1.49 WhCalculatedShown2024-03-27
GPT-3.5 TurboOpenAI113.41.56 WhDay-one estimateNone2024-01-25
Qwen3.7-PlusAlibaba147.31.60 WhInferredNot stated2026-06-02
DeepSeek-V4-FlashDeepSeek146.11.60 WhInferredShown2026-04-24
Qwen3-30B-A3B-Instruct (Jul 2025)Alibaba137.41.61 WhInferredShown2025-07-29
GPT-4.1 miniOpenAI135.01.61 WhDay-one estimateHidden2025-04-14
Mistral Large 2Mistral AI127.51.66 WhDay-one estimateShown2024-11-18
Qwen2.5-MaxAlibaba132.51.67 WhInferredNot stated2025-01-25
Qwen3-235B-A22B-Instruct (Jul 2025)Alibaba138.91.70 WhInferredShown2025-07-25
Gemini 3.1 Flash-LiteGoogle144.51.77 WhDay-one estimateHidden2026-03-03
DeepSeek LLM 67BDeepSeek110.51.81 WhCalculatedShown2023-11-29
Llama 3.3 70BMeta AI127.31.84 WhCalculatedShown2024-12-06
Llama 3.1-70BMeta AI125.91.84 WhCalculatedShown2024-07-23
Llama 3-70BMeta AI122.91.84 WhCalculatedShown2024-04-18
Llama 2-70BMeta AI113.81.84 WhCalculatedShown2023-07-18
DeepSeek-R1-Distill-Llama-70BDeepSeek-1.84 WhCalculatedShown2025-01-20
Hermes 2 Theta Llama-3 70BNous Research +1-1.84 WhCalculatedShown2024-06-20
Tulu 3 (Tülu 3) 70BAllen Institute for AI +1-1.84 WhCalculatedShown2024-11-21
Qwen1.5-72BAlibaba-1.86 WhCalculatedShown2024-02-04
Gemini 1.5 ProGoogle DeepMind126.91.86 WhInferredNone2024-05-14
Qwen2.5-72BAlibaba129.01.86 WhCalculatedShown2024-09-19
Qwen2-72BAlibaba125.31.86 WhCalculatedShown2024-06-07
Llama 3.2 90BMeta AI125.52.03 WhCalculatedShown2024-09-24
Mistral Medium 3.5Mistral AI141.52.16 WhDay-one estimateShown2026-04-28
Mistral Medium 3Mistral AI134.12.19 WhDay-one estimateNot stated2025-05-07
MiniMax-M2.7MiniMax145.92.34 WhInferredShown2026-03-18
Qwen 3.8 27BAlibaba151.62.41 WhInferredShown2026-08-14
DeepSeek-V3.1DeepSeek139.92.41 WhInferredShown2025-08-21
Qwen 3.6 PlusAlibaba147.52.44 WhInferredNot stated2026-03-31
Qwen3-30B-A3B-Thinking (Jul 2025)Alibaba139.62.45 WhInferredShown2025-07-30
Qwen 3.5 Plus (hosted 397B-A17B)Alibaba146.32.46 WhInferredNot stated2026-02-16
MiniMax-M2.5MiniMax146.72.51 WhInferredShown2026-02-12
WizardLM-2 8x22BMicrosoft-2.52 WhDay-one estimateShown2024-04-15
Gemini 2.0 FlashGoogle DeepMind +1134.72.53 WhInferredNot stated2025-02-05
Kimi K2 (Jul 2025)Moonshot140.12.56 WhInferredShown2025-07-12
Gemini 2.0 ProGoogle DeepMind135.12.57 WhInferredHidden2025-02-05
GPT-5.6 LunaOpenAI156.32.61 WhDay-one estimateHidden2026-07-09
Gemini 2.0 Flash ThinkingGoogle DeepMind +1135.42.65 WhInferredNot stated2025-01-21
GPT-4.5OpenAI136.72.67 WhInferredHidden2025-02-27
Qwen3-235B-A22BAlibaba139.32.69 WhCalculatedShown2025-04-29
Claude 3.5 SonnetAnthropic130.02.86 WhInferredNone2024-06-20
Claude 3 OpusAnthropic126.92.94 WhInferredNone2024-02-29
Claude InstantAnthropic120.32.94 WhInferredNone2023-08-09
Claude 2Anthropic120.23.06 WhInferredNone2023-07-11
Qwen 3.6 Max (Preview)Alibaba149.23.61 WhInferredNot stated2026-04-20
GPT-5.4 MiniOpenAI148.83.73 WhDay-one estimateHidden2026-03-17
DeepSeek-V3.2DeepSeek146.13.83 WhInferredShown2025-12-01
Grok 4 FastxAI144.23.91 WhInferredNot stated2025-09-19
DeepSeek-V3.2-ExpDeepSeek145.04.00 WhInferredShown2025-09-29
Claude Sonnet 4Anthropic141.74.18 WhInferredHidden2025-05-22
Grok-3 minixAI140.34.19 WhInferredNot stated2025-04-09
Inkling-SmallThinking Machines150.24.97 WhCalculatedShown2026-07-15
Mixtral 8x22BMistral AI122.05.13 WhCalculatedShown2024-04-17
Qwen3.5 397B-A17BAlibaba146.75.17 WhCalculatedShown2026-02-13
Llama 4 MaverickMeta AI132.25.17 WhCalculatedShown2025-04-05
Qwen3.8 Max (0902)Alibaba154.95.39 WhInferredNot stated2026-09-01
Muse Spark 1.1Meta AI154.05.60 WhInferredNot stated2026-07-09
Muse Spark 1.2Meta AI154.95.60 WhInferredNot stated2026-08-05
Claude Sonnet 5.5Anthropic164.45.64 WhDay-one estimateHidden2026-09-28
Claude Sonnet 5Anthropic156.45.78 WhDay-one estimateHidden2026-06-30
GLM-4.5Z.ai (Zhipu AI) +1-5.79 WhCalculatedShown2025-08-03
GLM-4.6Z.ai (Zhipu AI) +1-5.79 WhCalculatedShown2025-09-30
GPT-5.1OpenAI149.85.87 WhDay-one estimateHidden2025-11-13
Claude Opus 4.1Anthropic144.15.90 WhInferredHidden2025-08-05
Muse SparkMeta AI152.15.90 WhInferredNot stated2026-04-08
Qwen3-235B-A22B-Thinking (Jul 2025)Alibaba143.85.92 WhInferredShown2025-07-25
GPT-5OpenAI150.06.04 WhDay-one estimateHidden2025-08-07
Claude Opus 4Anthropic142.76.21 WhInferredHidden2025-05-22
GPT-4.1OpenAI136.86.21 WhDay-one estimateHidden2025-04-14
GPT-6 SolOpenAI162.66.44 WhDay-one estimateHidden2026-09-22
GPT-6.1 SolOpenAI166.36.64 WhDay-one estimateHidden2026-09-29
Claude 3.7 SonnetAnthropic141.16.66 WhInferredHidden2025-02-24
o1-miniOpenAI135.86.69 WhInferredHidden2024-09-12
o1-previewOpenAI134.86.70 WhDay-one estimateHidden2024-09-12
GPT-5.3 CodexOpenAI156.86.81 WhDay-one estimateHidden2026-02-05
GPT-5.2OpenAI153.56.83 WhDay-one estimateHidden2025-12-11
GPT-4oOpenAI128.76.92 WhDay-one estimateNone2024-08-06
Gemini 2.5 FlashGoogle DeepMind139.96.97 WhDay-one estimateHidden2025-06-17
GPT-5.6 TerraOpenAI159.47.22 WhDay-one estimateHidden2026-07-09
Claude Sonnet 4.5Anthropic146.87.42 WhDay-one estimateHidden2025-09-29
Claude Sonnet 4.6Anthropic152.47.48 WhDay-one estimateHidden2026-02-17
GPT-5.4OpenAI156.77.53 WhDay-one estimateHidden2026-03-05
Muse Spark 1.3Meta AI156.88.28 WhInferredNot stated2026-09-02
Qwen3.7-MaxAlibaba153.88.50 WhInferredNot stated2026-05-19
Llama 3.1-405BMeta AI128.810.2 WhDay-one estimateShown2024-07-23
GLM-5.2Z.ai (Zhipu AI)151.910.3 WhCalculatedShown2026-06-16
Gemini 3 FlashGoogle DeepMind151.810.9 WhDay-one estimateHidden2025-12-17
Mistral LargeMistral AI122.011.0 WhDay-one estimateNot stated2024-02-26
Grok 4.20xAI152.111.3 WhDay-one estimateNot stated2026-02-17
Grok 4.3 BetaxAI148.911.3 WhDay-one estimateNot stated2026-04-17
Kimi K2 ThinkingMoonshot146.011.6 WhCalculatedShown2025-11-06
Mistral Large 4Mistral AI146.911.9 WhDay-one estimateNot stated2026-10-06
DeepSeek-V3DeepSeek132.312.0 WhCalculatedShown2024-12-26
Nemotron 3 UltraNvidia145.813.5 WhCalculatedShown2026-06-04
GLM-5Z.ai (Zhipu AI)145.814.6 WhDay-one estimateShown2026-02-11
o3-proOpenAI147.415.7 WhInferredHidden2025-06-10
Gemini 3 ProGoogle DeepMind152.915.9 WhInferredHidden2025-11-18
DeepSeek V4 Flash 0731DeepSeek154.516.1 WhCalculatedShown2026-07-31
Qwen3-MaxAlibaba142.316.4 WhDay-one estimateNot stated2025-09-24
Claude Opus 4.8Anthropic158.316.7 WhDay-one estimateHidden2026-05-28
Claude Opus 5Anthropic162.617.0 WhDay-one estimateHidden2026-07-24
Claude Opus 5.5Anthropic166.817.0 WhDay-one estimateHidden2026-09-22
o3OpenAI146.917.0 WhInferredHidden2025-04-16
Gemini 3.6 FlashGoogle DeepMind154.217.2 WhDay-one estimateHidden2026-07-21
GLM-5.3-FlashZ.ai (Zhipu AI)153.217.6 WhCalculatedShown2026-08-20
Gemini 3.8 FlashGoogle DeepMind157.217.6 WhDay-one estimateHidden2026-09-02
Gemini 3.7 FlashGoogle DeepMind157.217.7 WhDay-one estimateHidden2026-08-13
Claude Opus 4.7Anthropic156.318.2 WhDay-one estimateHidden2026-04-16
Claude Opus 4.6Anthropic155.318.7 WhDay-one estimateHidden2026-02-05
Claude Opus 4.5Anthropic150.219.1 WhDay-one estimateHidden2025-11-24
GLM-4.7Z.ai (Zhipu AI)143.819.1 WhCalculatedShown2025-12-22
Gemini 3.5 FlashGoogle DeepMind154.419.7 WhDay-one estimateHidden2026-05-19
GPT-5.5OpenAI158.921.1 WhDay-one estimateHidden2026-04-23
GPT-5.6 SolOpenAI161.721.5 WhDay-one estimateHidden2026-07-09
MiniMax-M3MiniMax147.122.0 WhCalculatedShown2026-06-01
GPT-4 Turbo (Nov 2023)OpenAI126.523.2 WhDay-one estimateNone2024-01-25
DeepSeek V4 Pro 0813DeepSeek155.325.9 WhCalculatedShown2026-08-13
DeepSeek-V4-ProDeepSeek149.225.9 WhCalculatedShown2026-04-24
GLM-5.1Z.ai (Zhipu AI)150.027.6 WhDay-one estimateShown2026-04-07
Kimi K2.6Moonshot150.932.9 WhDay-one estimateShown2026-04-20
Kimi K2.7 CodeMoonshot149.932.9 WhDay-one estimateShown2026-06-12
Kimi K2.5Moonshot148.032.9 WhDay-one estimateShown2026-01-27
Gemini 4 ArgonGoogle DeepMind164.634.7 WhInferredHidden2026-09-30
Grok 4.6xAI156.339.8 WhDay-one estimateNot stated2026-08-12
Grok 4.5xAI153.939.8 WhDay-one estimateNot stated2026-07-08
Gemini 3.1 ProGoogle DeepMind154.749.7 WhDay-one estimateHidden2026-02-19
DeepSeek V4.1 FlashDeepSeek155.654.2 WhCalculatedShown2026-09-09
Claude Fable 5Anthropic162.154.5 WhDay-one estimateHidden2026-06-09
Claude Fable 5.1Anthropic164.754.9 WhDay-one estimateHidden2026-09-01
DeepSeek-R1DeepSeek139.063.9 WhCalculatedShown2025-01-20
GPT-6 AstraOpenAI166.464.0 WhDay-one estimateHidden2026-09-03
Qwen 3.8 MaxAlibaba156.966.9 WhCalculatedNot stated2026-08-02
GPT-4 (Mar 2023)OpenAI123.267.5 WhDay-one estimateNone2023-03-14
Kimi K3Moonshot157.469.9 WhCalculatedShown2026-07-16
GLM-5.3Z.ai (Zhipu AI)155.969.9 WhCalculatedShown2026-08-14
Gemini 2.5 Pro (Mar 2025)Google DeepMind142.472.1 WhDay-one estimateHidden2025-03-31
InklingThinking Machines148.688.8 WhCalculatedShown2026-07-15
Grok 4xAI146.4258 WhDay-one estimateNot stated2025-07-09
Grok 3xAI138.3258 WhDay-one estimateNot stated2025-04-09
GPT-5 ProOpenAI150.7265 WhDay-one estimateHidden2025-10-07
GPT-5.2 ProOpenAI156.0670 WhDay-one estimateHidden2025-12-11
GPT-5.5 ProOpenAI162.0730 WhDay-one estimateHidden2026-04-23
GPT-5.4 ProOpenAI159.0796 WhDay-one estimateHidden2026-03-05
AI Co-MathematicianGoogle DeepMind--No dataHidden2026-05-08
Gemini 2.5 Deep ThinkGoogle +1--No dataNot stated2025-08-01
Grok 4 HeavyxAI--No dataNot stated2025-07-10
Muse Spark OpenMeta AI--No dataShown2026-10-15
Muse WonderMeta AI--No dataNot stated2026-10-12
QWQ-PlusAlibaba--No dataNot stated2025-04-08
Qwen PlusAlibaba--No dataNot stated2025-04-28
Qwen-TurboAlibaba--No dataNot stated2024-11-01
What we will and won't claim. Our figures are accurate to roughly a factor of two to three when checked against laboratory measurements. So we only state a difference when it is at least 5×, comfortably outside that margin. Where two models are closer than that, we say they are too close to call rather than inventing a ranking.

Capability is Epoch AI's Epoch AI / Capabilities Index, a single score combining 58 separate tests, coding, using tools, holding a long document in mind, writing and science reasoning among them. It runs from roughly 55 to 170. We use the composite rather than any one test because a model can be excellent at science questions and mediocre at everything else. A handful of older models have no index score; we leave their capability blank rather than filling it from a different test, which would put them in the wrong place in the order.

Energy is calculated, not metered, no AI company publishes per-answer energy. We use the EcoLogits model, which works from how much of a model fires per word and how many words it writes: a 7,000-word-ish answer for a complex prompt, 300 for everyday chat, on a world-average electricity grid.

Calibration. The raw formula assumes a model has hardware largely to itself, which overstates real energy use. We corrected it against 46 samples vs ML.ENERGY on B200, batch 32-128, applying a factor of 0.53518.

Every model is calculated the same way, including those we have lab measurements for. Measured figures swing eightfold with how many requests a server handles at once, and only exist for open models, so using them directly would rank serving setups, not models. They are used to check our work instead.

Day-one estimates are models whose makers have never published a size. We bracket it from comparable models, so those bars are wide on purpose. The largest figures on this page carry the most uncertainty.

Hidden thinking. 65 of the models here generate reasoning text you never see. OpenAI, Anthropic and Google all confirm they bill for it while returning only a summary, or nothing. Our figures count a fixed answer length, so for those models the number shown is an understatement, not an estimate. Each one is marked in the table.

Sources. Epoch AI (capability, model sizes, roster), EcoLogits (energy model), ML.ENERGY (laboratory measurements). All public, all free.

Reuse. These figures are published under CC BY 4.0, use them anywhere, including commercially, as long as you credit The AI Footprint and link back. Read the full methodology →

Privacy. We count page visits in aggregate and set no cookies. What that means →