Every comment is a model call on your own key. On a single request the price is invisible; at ten thousand a month it becomes a line item you can see.
What you are paying for
Two things: what you send and what the model returns. You send the prompt plus the post; the model returns a short comment. So most of the spend sits on the input side and grows with prompt length, not with the answer.
The practical consequence: a two-page prompt costs more than a ten-line one for the same result. Detailed instructions are worth having, but every extra line gets multiplied by the number of posts in a month.
Choosing a model by cost
The gap between light and heavy models is orders of magnitude, not percent. For a two-line comment a heavy model is usually overkill: it writes better, but not better enough to justify the difference at volume.
Sensible order: start light, read a hundred real comments in your niche, and only move up if the quality genuinely fails. Doing it the other way round costs more and buys nothing.
What else moves the bill
- Filtering. A post that will never receive a comment should not reach the model — that call is already paid for.
- Duplicates. Processing the same post twice bills you twice.
- Your own router. Connecting an OpenAI-compatible router gives you provider pricing without a middleman markup.
Estimating a month up front
Multiply expected daily comments by thirty and by the per-call price of your chosen model. Compare with the subscription: model spend is usually comparable or larger, which is fine — but better known in advance. How it folds into cost per lead is in the lead cost piece.