It is per question.

The models "cost" is based on the number of "tokens" they burn to "reason". So many AI providers charge a $$/million tokens of input and a $$/million tokens of output. That dollars per tokens is different for different models. More advanced models have higher pricing. This pricing is often obscured by AI pltdorms with various monthly subscriptions for unknown amounts of query as well.

I did my best to convert the costs I saw, into sats per query for users. The way the payments work is by loading wallet via lightning or cashu on chat.routstr.com . The cashu tokens are sent to my proxy per query, then the proxy passes the query to the ai api keys, refunds the difference in estimated costs and actual costs and gives you your chat response.

Every question is an individual cashu spend. Some queries are 1 sat while others can be 5,000+ sats depending on model and action. This is the same with fiat ai platforms that charge per query where bigger computation or long conversations become heavier total outputs.

I will be playing heavily with this over the week and tweak things into line but I need it live to see how it interacts as a proxy on both ends.