How are you pricing your AI product, metered or flat?

We skipped the wallet/token pricing model everyone recommends for AI apps. Went with a flat subscription instead: free roast to hook people, paywall on depth (history, category comparisons). The token model mostly exists to shift unpredictable cost risk onto the user; our per-user cost is low enough that we don't have that problem, so it felt like solving a risk we don't actually carry.

Also just wired up real per-user cost tracking (Helicone) instead of eyeballing it. Curious how many of you actually know your margin per customer vs. guessing.

How did you land on your model? Did anyone else skip the token/wallet route for a similar reason, or is there a case for it I'm missing?