[daxfohl]

Makes sense, and each model has a max context length, so they could charge per token assuming full context by model if they wanted to assume worst case.