How are you pricing AI features without getting burned by token costs？
I’m curious how other AI tool builders are handling pricing.
A lot of AI products start with a simple plan:
- free trial
- $19/mo
- maybe a higher tier later
But once users actually start using the product, the cost structure gets weird fast.
One user might send 20 short prompts a day.
Another user might upload long documents, trigger agents, retry failed outputs, or run workflows in the background.
Same subscription price, completely different model cost.
The tricky part is that users usually do not want to think in tokens. They think in outcomes:
- generate this report
- summarize this file
- create this image
- analyze this dataset
- run this agent workflow
But under the hood, every workflow has a different cost profile.
This is something we are thinking about at EvoLink Model Router: how to help AI products route tasks across models, compare cost/performance, and keep usage from quietly eating the margin.
Curious how you are handling this:
- Do you charge users by seat, usage, credits, or fixed plans?
- Do you expose token/credit usage to users?
- Have you ever had a “good” user become unprofitable because they used the AI feature too much?
- At what point did you add limits, routing, or cheaper fallback models?
The comment from aryan_sinh nails the core tension: pricing that looks clean on the page breaks when power users show up. What worked for us was separating delivery cost from perceived value. Cheap operations like short prompts get bundled into the base plan. Expensive workflows like document analysis and multi-step agent runs go into a usage pool. We also route across models based on task complexity cheap local models for simple lookups, frontier models for reasoning-heavy work. The routing logic is invisible to the user so they still get the outcome without us eating the margin. The real unlock was measuring cost-per-outcome instead of cost-per-token.
I think the dangerous part is that most AI pricing starts from what feels simple to sell, not from where margin can quietly break.
Users don’t think in tokens, but founders also can’t price every workflow like it costs the same to serve.
So the hard decision is not just seat vs credits vs usage. It is deciding which outcomes should feel unlimited, which ones need soft limits, and which expensive workflows need routing/fallback before they become margin leaks.
That pricing layer probably has to be designed around behavior, not just model cost.
Happy to put the tighter pricing structure in writing if useful. This is one of those areas where a simple $19 plan can look clean on the page and still break once power users show up.