Model sheets
The footprint of every model, per million generated tokens
The TokenClimate registry tracks 26 models from six providers. Per million generated tokens, the range goes from 23 gCO₂e for GPT-6 Luna, the most frugal, to 29.5 kgCO₂e for DeepSeek V4 Pro, a ×1,261 spread.
Every factor is published with its source and confidence level, updated on . These figures also power the calculator and the full method is in the methodology.
| Sheet | |||||
|---|---|---|---|---|---|
| GPT-6 Luna | OpenAI | 23 g | 274 mL | low | sheet |
| Mistral Small 4 | Mistral | 24 g | 919 mL | medium | sheet |
| GPT-4o mini | OpenAI | 31 g | 367 mL | low | sheet |
| Mistral Medium 3.5 | Mistral | 58 g | 2.3 L | medium | sheet |
| GPT-5.4 nano | OpenAI | 62 g | 729 mL | low | sheet |
| GPT-5.6 Luna | OpenAI | 62 g | 729 mL | low | sheet |
| GPT-4.1 mini | OpenAI | 145 g | 1.7 L | low | sheet |
| GPT-5.4 mini | OpenAI | 194 g | 2.3 L | low | sheet |
| Mistral Large 3 | Mistral | 290 g | 11.3 L | medium | sheet |
| Claude Haiku | Anthropic | 475 g | 7.5 L | medium | sheet |
| GPT-6 Sol | OpenAI | 661 g | 7.8 L | low | sheet |
| GPT-6.1 Sol | OpenAI | 661 g | 7.8 L | low | sheet |
| GPT-4.1 | OpenAI | 677 g | 7.9 L | low | sheet |
| Gemini 2.5 Flash | 685 g | 7.3 L | low | sheet | |
| GPT-4o | OpenAI | 739 g | 8.7 L | low | sheet |
| GPT-5.4 | OpenAI | 779 g | 9.1 L | low | sheet |
| GPT-5.6 Terra | OpenAI | 779 g | 9.1 L | low | sheet |
| Claude Sonnet | Anthropic | 950 g | 15.0 L | high | sheet |
| Claude Opus | Anthropic | 1.9 kg | 30.0 L | medium | sheet |
| GPT-5.5 | OpenAI | 2.4 kg | 27.8 L | low | sheet |
| GPT-5.6 Sol | OpenAI | 2.4 kg | 27.8 L | low | sheet |
| Claude Fable | Anthropic | 3.8 kg | 60.0 L | low | sheet |
| GPT-6 Astra | OpenAI | 4.3 kg | 50.7 L | low | sheet |
| Gemini 2.5 Pro | 7.5 kg | 80.5 L | low | sheet | |
| DeepSeek V4 Flash | DeepSeek | 7.9 kg | 85.8 L | low | sheet |
| DeepSeek V4 Pro | DeepSeek | 29.5 kg | 321 L | low | sheet |
The table counts generated tokens. An input token weighs much less, and real usage mixes input, cache and output: the calculator works out your own split.
Retired or replaced models (14)
Retired, or being retired, by their provider, with the shutdown date when one is published. The sheets stay online for reference and to quantify past usage.
- DeepSeek Chat (V3) (shutdown on )
- DeepSeek Reasoner (R1) (shutdown on )
- Gemini 1.5 Flash
- Gemini 1.5 Pro
- Gemini 2.0 Flash (shutdown on )
- Llama 3.1 8B (shutdown on )
- Llama 3.3 70B (shutdown on )
- Mistral Large 2 (shutdown on )
- GPT-4 Turbo (shutdown on )
- GPT-4.1 nano (shutdown on )
- o1 (shutdown on )
- o3 (shutdown on )
- o3-mini (shutdown on )
- o4-mini (shutdown on )