The number that decides whether your AI feature works isn’t the total API bill — it’s the cost to serve one user. Compute it by multiplying a typical request’s tokens by the model’s rates, then by how many requests an average user makes a month. If that figure sits comfortably below what you charge, you have a business; if it doesn’t, you’re paying people to use your product. The AI cost per user calculator builds this from the ground up and shows your margin and break-even price.
Why per-user cost, not total cost
Founders watch the monthly API invoice and panic or relax based on the total. That’s the wrong signal. A rising total can mean healthy growth if each new user is profitable, and a low total can hide a feature that loses money on every heavy user. Unit economics — cost and margin per user — is the number that tells you whether growth helps or hurts.
It’s also the number you can act on. You can’t easily shrink a total bill without losing customers, but you can cut per-user cost with caching, cheaper models and usage caps, and you can raise price above break-even. Everything useful happens at the per-user level.
The formula
Cost per user has three inputs:
- Tokens per request — input (the prompt, context and any retrieved documents) plus output (what the model generates).
- Model rates — the per-token price for input and output, which differ, often by 3 to 5 times.
- Requests per user per month — how often an average user triggers the feature.
Cost per request = (input tokens × input rate) + (output tokens × output rate). Multiply by monthly requests and you have the per-user cost. Subtract it from your price to get margin.
A worked example
Say you’re shipping an AI summariser. An average user runs it 40 times a month. Each request sends 3,000 input tokens (the document plus your prompt) and gets back 500 output tokens.
On a mid-tier model priced around $3 per million input tokens and $15 per million output:
- Input: 3,000 × $3 / 1,000,000 = $0.009
- Output: 500 × $15 / 1,000,000 = $0.0075
- Per request: ~$0.0165
- Per user per month: 40 × $0.0165 = ~$0.66
Charge $10 a month and your AI cost is $0.66 — a healthy margin before hosting and support. But watch what a power user does: someone running it 500 times a month costs ~$8.25, which nearly wipes out that $10 on AI alone. Your average looks fine; your heaviest user is almost underwater. That gap is the whole game.
Protecting margin against power users
Averages lie because AI usage is lopsided — a small fraction of users generate most of the requests. Price for the average and your heavy users quietly erase the profit from everyone else. Three defences:
- Cap or meter. Include a generous monthly allowance, then meter or throttle beyond it. Most users never hit the cap; the few who do cover their own cost.
- Tier by usage. Move heavy users to a higher plan whose price sits above their cost.
- Cut cost per request. Cache the stable parts of your prompt, trim retrieved context to what’s needed, and route simple requests to a cheaper model. Switching models alone can halve the bill — how much you save switching AI models shows the math.
If your feature is agent-shaped — multiple back-and-forth turns per task — cost climbs faster than a single request suggests, because each turn resends the whole conversation. Why AI agents cost so much explains that compounding, which matters a lot when you’re estimating requests per user.
Set your price above break-even
Break-even is the price where what you charge equals the AI cost of serving that user. It’s the floor, not the target — you still have hosting, retrieval, moderation and support to cover on top. The AI cost per user calculator shows your break-even alongside the per-user cost and flags a negative margin, so you can price the feature before the invoice teaches you the hard way. Remember it’s the model-API floor only; layer the rest of your infrastructure on top before you commit to a number.
