I wouldn't call input a vast majority of token usage. In my experience output is 25%-33% of the cost.