GPT-4o: cost, energy and CO₂ per token
- Provider
- OpenAI
- Confidence
- low
GPT-4o emits about 702 g of CO₂e, uses 8.2 L of water and 1.6 kWh of electricity per million generated tokens.
Key figures, per million tokens
public price, in eurosOn the bill
GPT-4o1M generated tokens€8.69
1M input tokens€2.17
1M tokens read from cache€1.09
Not on the bill
CO₂e, 1M generated tokensabout 5 km in a combustion car702 g
Water8.2 L
Energy1.6 kWh
1M input tokens117 g, 1.4 L, 265 Wh
GPT-4o among the 34 models in the registry
GPT-4o emits 5% more than the median of the 34 models in the TokenClimate registry (668 g per million generated tokens), that is ×36 compared with Mistral Small 4, the most frugal.
Sources and method
Full equations and assumptions in the methodology.
The calculation, per million generated tokens
1,591 Wh × (1.12 × 0.35 g/Wh + 0.049 g/Wh) = 702 g
- Energy per token
- The per-token energy factor behind the figures above, published with its citation. EcoLogits v0.11.3
- Infrastructure parameters
- Datacenter parameters applied to this provider (the equations live in the methodology). Jegham et al., arXiv:2505.09598 (v6)
PUE 1.12 Grid carbon intensity (CIF) 0.35 kgCO₂e/kWh On-site water (WUE) 0.3 L/kWh Off-site water (EWIF) 4.35 L/kWh Amortised manufacturing (embodied) 49 gCO₂e/kWh - Price
- Public provider prices. developers.openai.com, pricing
Frequently asked questions
- How much CO₂e for 1 million tokens with GPT-4o?
- About 702 g per million generated tokens and 117 g per million input tokens (factors ecologits-0.11.3-2026-10 of the TokenClimate registry).
- Where do these figures come from?
- Factors published in TokenClimate's versioned registry, each with its citation on this page. The full equations and parameters are in the public methodology, at tokenclimate.com/en/methodology.
- What is the confidence level?
- Low confidence: extrapolated from another model, or estimated by a parametric tool.
- How can you measure your organisation's real usage?
- The TokenClimate dashboard connects to your organisation's Anthropic, OpenAI or OpenRouter admin keys (stored encrypted) and measures cost and CO₂e per team and per model. You can see it in the demo: tokenclimate.com/demo. Locally, the open source plugin claude-carbon measures a Claude Code session.
Does your organisation use GPT-4o?
The dashboard pulls your volumes every day from your Anthropic, OpenAI or OpenRouter admin keys, and puts cost and CO₂e on each team and each model.