Here’s how to price AI so your costs don’t explode

If your product uses AI, every click increases your costs.

Here's how to turn LLM pricing into plans, credits, and limits that won’t come back to bite you.

1. Forget tokens — think in “AI actions”

If you start from “tokens”, your brain melts. Start from what your user actually does.

Write down the main things in your product that use AI. For example:

• “Ask my assistant a question”
• “Summarize this document”
• “Rewrite this text in a new tone”
• “Run an agent over these items” (e.g., emails, leads, notes)

Each of these is a single AI action. You’ll want to break every workflow or feature into single actions.

2. Separate cheap models from expensive models

In your product, you will usually have two kinds of models:

• Cheap models – fast, good enough for simple tasks, low cost per call
• Expensive models – better reasoning, handle more context, much higher cost per call

Give them simple labels in your system:

• Standard AI → cheap mode
• Deep AI → expensive model

These labels are only for you and your team. Users don’t choose them.

From here on, we’ll talk about Standard vs Deep, not specific model names. That way this works no matter which AI provider you use.

3. Work out what each action costs

Use the same action list from step 1.

Now you want to know: “How much does one of these actions cost me, roughly?”

If you don’t have users yet

You can’t get real prices, but you can measure size.

For each action:

• Input: short / medium / long?
• Output: short / medium / long?
• Calls: one call or many calls (agent)?

Use this:

• Short = 1–2 sentences, or a short chat
• Medium = a few paragraphs, about one page
• Long = many pages, or a long chat history

Examples:

• Chat reply → short in, short out, 1 call → very cheap
• Summarize 3 pages → long in, short out → cheap
• Summarize a long report → very long in → medium
• Agent over ~100 items (e.g. 100 emails, 100 leads, 100 notes) → many calls → expensive

Now you know which actions are light and which ones are heavy.

If your product is live

Now you can get real numbers.

1. Track the action name For each AI call, try to record which feature it came from (e.g. “chat_reply”, “short_summary”, “agent_run”).
2. Look at one action at a time In your logs or analytics, filter by that action name and find:
   • total cost for that action
   • number of runs
3. Compute average cost

Cost per action ≈ total cost for that action ÷ number of runs

Do this for every action on your list.

At the end of this step you have:

• A list of all actions
• The average cost of one run of each action

You’ll use this in the next step to get cost per user per month.

4. Turn action costs into cost per user

You now know the cost of one run of each action.

Now ask: “For a normal paying user, how many times does each action run in a month?”

If you have users: use analytics to see how often each action is used per user.

If you don’t: make a simple and honest estimate for each action.

For each action: Cost per user for this action = (uses per month) × (cost per action)

Add all actions together: AI cost per active user per month.

You will use this number to:

• Choose your prices
• Set your credits and limits

5. Add your other costs and choose a safe margin

AI is not your only cost.

You also pay for things like:

• Servers
• Storage
• Support
• Tools and fees

If you have users:

• Take all these non-AI costs for one month
• Divide by the number of active paying users

That gives other cost per user.

If you are pre-launch: Make a simple estimate, and update it later when you have real data.

Then: Total cost per user = AI cost per user + other cost per user

Later, you’ll set your prices so they are well above this total cost.

Remember: As you grow, this number can change (for example, servers can get cheaper per user).

So check it again from time to time.

6. Turn your costs into credits and limits

You can charge in many ways:

• Just a subscription
• Subscription + included credits
• Or pure usage (pay per use)

Here we’ll assume you use credits, either alone or on top of a subscription.

Give each action a credit cost

Use your cost per action from step 3.

Pick simple numbers like 1, 2, 5, 10 and assign them to actions:

• 1 credit → one chat reply
• 2 credits → one short document summary
• 5 credits → one deep analysis of a long document
• 10 credits → one agent run over ~100 items

More expensive actions should use more credits.

Decide how many credits are in a plan

From steps 4 and 5, you know:

• How much AI you spend on a normal user
• And your total cost per user

Decide how much AI cost you’re happy to include in this plan (for example, “about $5 of AI per user”).

Then:

• Pick a number of credits so that a normal user can do their usual work
• And your AI cost stays around that target

In the UI, you can show something simple like: “You’ve used 700 / 1,500 AI credits this month.”

7. Make the product waste less AI by design

You can also save a lot by changing how you call the model.

Here are a few simple tricks that don’t hurt UX:

• Use the cheap model most of the time. Use it for normal chat, small edits, and short summaries. Most people won’t see the difference. They only care that it’s fast and clear.

• Keep context small. Don’t send the whole chat every time. Make a short summary of old messages and send that instead. For documents, ask the user what they care about, and only send those parts.

• Keep answers short. In your prompt, say the AI something like: “Keep the answer under 200 words.” “Give 3 bullet points only.”

• Cache old answers. If people ask the same question many times, save the answer. Next time the question is the same, show the saved answer instead of calling the AI again.

These simple steps can cut your AI bill a lot.