AI Cost per User: How to Price an App Built on an LLM

Jonhisking · October 2, 2026

If you are building a product on top of an AI model, the most important number in your business plan is not the price per million tokens. It is AI cost per user per month. That number decides whether a free tier is affordable, what you can charge, and whether your heaviest users make or lose you money. Here is how to estimate it.

Step 1: describe a typical user

Start with behavior, not tokens. For a chat-style assistant, for example:

• Active on 15 days a month
• 10 messages on each active day, so 150 messages a month
• Each message sends about 2,500 input tokens (system prompt, recent conversation, the new message)
• Each answer is about 400 output tokens

Measure these from a prototype if you can. Paste a real system prompt and a real conversation into the token counter rather than guessing.

Step 2: cost per message

At API prices checked October 1, 2026:

Step 3: cost per user per month

Multiply by 150 messages:

On a $9.99 monthly subscription, even the most expensive model leaves room for a profit on a typical user. But typical users are not the problem.

Step 4: plan for the heavy tail

Usage in AI products is very uneven. A small share of users often accounts for a large share of messages. Suppose the top 5% of users send ten times as much as the typical user, with longer conversations that double input size:

• 1,500 messages a month at 5,000 input and 400 output tokens
• On GPT-6 Sol: 1,500 × ($0.010 + $0.004) = $21 a month

On a $9.99 plan, that user loses you money every month. This is why most AI products have usage limits, even on paid plans.

Step 5: price the free tier

Free users cost money too. If free users average 30 messages a month on Gemini 3.8 Flash, each costs about $0.10. With 10,000 free users that is around $1,000 a month, before any of them pay. Decide what conversion rate makes that worthwhile.

Levers that change the numbers

1. Model routing. Send most messages to a small model and only hard ones to a large model. If 80% of messages can go to a model a tenth of the price, the blended cost drops by roughly 70%.
2. Prompt caching. If your system prompt is long, caching can cut the cost of that part by around 90%. See Prompt caching explained.
3. History limits. Summarize old turns instead of resending whole conversations. Input per message stops growing.
4. Output limits. Shorter default answers, with "tell me more" for people who want detail.
5. Usage caps per plan. Generous enough for typical users, firm enough to protect against the heavy tail.
6. Language. If most of your users write in Japanese or Korean, expect 1.4–1.8 times the tokens of English for the same conversations.

A simple model to start from

monthly cost per user = messages per month
                      × (input tokens × input price + output tokens × output price) / 1,000,000

Calculate it for a typical user, a heavy user and a free user, and check each one against what that user pays. Then revisit the numbers every few months: model prices tend to fall, and your users' behavior will change.

Related

• How to estimate your AI API bill before you build
• ChatGPT Plus or the API: which is cheaper?
• Why output tokens cost more