Adding an AI chat or assistant to a product is easy. Knowing what it will cost per customer, and whether your price covers it, is harder. Token prices are quoted per million, which makes them look negligible until real usage arrives.
Think in cost per active user
Your customers pay per month, so your AI cost should be measured the same way: the average monthly cost of one active user. It depends on how many conversations they have, how long each one runs, and the model you use.
Don't forget chat history
In a normal chat, every new message re-sends the whole conversation so far. The input grows with every turn, so a 12-message conversation costs far more than two 6-message ones.
Input tokens
6 × 800 + 6 × 60 + 360 × 6 × 5 ÷ 2equals10,560
Output tokens
6 × 300equals1,800
Per conversation
10,560 ÷ 1M × $2.00 + 1,800 ÷ 1M × $10.00equals$0.0391
Per month
$0.0391 × 10,000 conversationsequals$391.20
That is $0.39 per active user per month. If the feature is part of a $10 plan, AI costs take about 4% of revenue, which is comfortable. If your heaviest users chat ten times as much, their cost is ten times higher, so look at your usage distribution, not just the average.
Turning cost into a price
- Set a target gross margin for the feature, for example 80%. Price ≥ cost per user ÷ (1 − 0.80).
- Put limits on the heaviest usage: message caps, a cheaper model for simple requests, or a higher tier for power users.
- Recheck when models change. Prices fall regularly, and a model switch can halve your cost overnight.
Tip: Long system prompts and documents are re-sent with every message. Prompt caching, offered by most providers, can cut that repeated input cost sharply.