Small numbers. Full context.
Recorded responses
Modeled inference electricity
Illustrative CO₂e scenario difference
These counters summarize responses recorded in this installation. They are not independently audited. Energy is modeled from token usage, not measured at the data center. The carbon comparison is a hypothetical scenario, not verified avoided emissions or a carbon offset. Development or demonstration usage may be included.
Show the math.
Name the assumptions.
01 / Electricity
Estimated watt-hours = (output tokens + 0.1 × input tokens) ÷ 1,000 × the selected model’s rate. Rates include an assumed data center overhead (PUE of approximately 1.3).
Mistral Small 3.2: 0.1 Wh per 1,000 output tokens; DeepSeek V4 Flash: 0.3 Wh per 1,000 output tokens; Gemma 4 26B: 0.1 Wh per 1,000 output tokens; GPT-OSS 120B: 0.3 Wh per 1,000 output tokens; Llama 3.3 70B: 0.5 Wh per 1,000 output tokens; Qwen3 235B: 0.6 Wh per 1,000 output tokens.
These are provisional modeling assumptions. Actual hardware, batching and request length can change consumption substantially.
02 / Carbon scenarios
The same modeled electricity is multiplied by two illustrative factors: 15 gCO₂e/kWh for a renewable electricity scenario and 390 gCO₂e/kWh for a US grid scenario. The card shows the difference.
These fixed assumptions are not live grid data or measured provider emission factors. They cannot establish the emissions avoided by your conversation.
03 / What’s outside the numbers
Model training, hardware manufacturing, water use, network traffic, your device and the website’s own infrastructure are not included. This is not a complete lifecycle assessment.
04 / What you can check
Provider disclosures are linked on our approach page. As better measurements become available, the model can improve. A useful estimate should always travel with its limitations.
Explore provider disclosures →