I am using Claude Professional, and have a usage counter (Claude Code Usage VSCode plugin) installed. It tells me that I spend tokens worth $80 per day, but I pay only $20 per month.
Where is this this huge difference coming from?
This is the huge difference between API costs and subscription costs of many providers (actually chinese ones are more honest/less subsidized). Judging by looking electricity consumption and depreciation of hardware I really think they are strongly subsidizing tokens currently for subscriptions, but, also, at the same time they are also making huge margins with API prices. I guess they would break even by adding a 0 on the right to all subscriptions 
but this is purely speculative. I guess their strategy is either to increase slowly the costs of subscriptions, or to reduce the number of tokens for a given subscription. Not that I think this works since contenders/alternatives are everywhere (fortunately). There is also the possibility that smaller models are enough for many tasks and those are maybe more economically viable, but we don’t know the size of closed models so we can only judge from the capabilities of open source models of a given size.
AI development is heavily subsidized: venture capital, other business revenues, public funding, borrowing, etc. That daily $80 is based on estimates that likely only cover immediate inference costs, not the operational overheads, infrastructure, and research that made it possible.
