If you’re running an AI startup, billing is a distraction. Pricing by API calls? GPU usage? Tokens? Many founders I’ve spoken to struggle with figuring out the best model—some even lose money because of unpredictable costs.
We’ve seen AI startups:
- Overpaying for compute-heavy workloads
- Struggling with OpenAI API pricing & limits
- Finding it hard to offer usage-based pricing to customers
Would love to hear how you’re solving this. Have you built an internal system, or are you using a solution? What’s the biggest challenge for you?
2 Likes
For us, we ended up flat fee subscription+ volume of tokens included; and separate billing per preset token number of overload (when included tokens are all consumed).
2 Likes
Interesting! So a friend of mine runs a AI startup too, and she and her team struggled to accurately measure their internal API cost “per action”. Then ended up with a fixed cost by taking some averages. They don’t like this, but the feedback that I got was that the other tools/APIs that they use have varied costs, behind the scenes. Atleast in your case, are you also averaging a part of cost behind the token or were you successfully able to calculate your internal cost of APIs?
Our subscription model offers x amount of text pages for legal document analysis. We use a common practice of fixed number of characters per source text page: 1800.
So based on average numbers of token we spend per page processing we came up with a fixed dollar amount per page.
Low volume subscription have bigger dollar amount per page, high volumes subscriptions have lower amount per page.
Stripe products are set up to track usage of characters converting them to pages, we also track usage of tokens for our internal statistics, basically to keep an eye on spent amount.
When the user exceeds the number of pages included in the subscription, we have separate pricing for additional pages which is usually cheaper. Depending on the billing configuration we either accept additional pages or require them to add balance to their account.
We’ve chosen going with approximation of our costs to simplify the billing and calculations on our end and for the customer to understand the billing structure.
The thing is that basically a page processing costs me around 3 cents, and the average selling price is around 4 USD. So with those margins we’re actually don’t care if there are some tokens left on either side. We just keep an eye that our approximation stays more or less accurate.
1 Like
This is super insightful! Thanks @sergeliatko for sharing. I’ll be sharing this with a few friends.
2 Likes