This talk at Green IO 2025 is really important. It demonstrates that the energy consumption of #genAI tools is much higher (1000x for small queries) than an estimate based on the input or output tokens.
This is because the tool does not simply send your prompt to the model. It adds a system prompt which is very large, and based on key words in the input, system prompt and generated output, invokes other LLM-based tools. (1/2)