# How to Price an AI Product When Every User Costs You Different Money

> How to price an AI product when each user costs you different money: measure cost per user, set a margin floor, pick a plan shape and write honest limits.

---
url: "https://earlyhunt.com/studio/how-to-price-an-ai-product-when-costs-vary-per-user"
markdown: "https://earlyhunt.com/studio/how-to-price-an-ai-product-when-costs-vary-per-user.md"
type: studio_post
title: How to Price an AI Product When Every User Costs You Different Money
slug: how-to-price-an-ai-product-when-costs-vary-per-user
published: "2026-10-05T04:55:58.463Z"
updated: "2026-10-05T04:56:03.718Z"
---

## Summary

How to price an AI product when each user costs you different money: measure cost per user, set a margin floor, pick a plan shape and write honest limits.

## Article

Picture two customers paying you the same $19 a month. One costs you forty cents in model fees. The other costs you thirty-six dollars. Normal SaaS doesn't work like that: a project-management seat costs about the same to serve whether someone logs in once a week or all day. An AI tool spends money every time someone clicks "generate," and some people click it all day.

Price it like normal SaaS and your lightest users quietly pay for your heaviest ones. This guide covers how to price anyway: measure what each user really costs, set a margin floor, pick a pricing shape, write limits that don't feel like a trap, and handle the few people who use ten times more than everyone else.

This is the cost side of AI pricing. If you haven't picked a starting price at all, read [how to price your first SaaS when you have almost no data](https://auraplusplus.com/studio/how-to-price-your-first-saas-when-you-have-almost-no-data) first. Here we assume you have a rough price and want to make sure it doesn't lose money.

## Why flat pricing breaks for AI tools

Most model providers charge per token, which is roughly a chunk of a word. [OpenAI's API pricing page](https://developers.openai.com/api/docs/pricing) and [Anthropic's pricing docs](https://platform.claude.com/docs/en/about-claude/pricing) both list prices per million tokens, with separate rates for input and output, and output costs more than input on both. They also list cheaper rates for cached input and for batch jobs, plus separate charges for some tools, like web search.

So your cost per user depends on things your pricing page never mentions:

- How often they use the product
- How long their inputs are (a pasted 40-page PDF vs a one-line question)
- How long the outputs are
- Which model handles the request
- How many retries, regenerations and background jobs you run on their behalf

A flat plan treats all of that as one number. Fine with twenty users and a small bill. Not fine the week someone wires your tool into a script and runs it overnight.

## Step 1: Measure the real cost per active user

You don't need a data warehouse. A spreadsheet and last month's bills are enough.

### Pull three things

1. **Your provider bills** for the last full month, for every AI service you use: the main model, embeddings, transcription, image generation, search tools.
2. **Your request logs, tagged by user.** Most model APIs return token counts with each response, usually in a usage field. If you aren't saving those counts next to a user ID yet, start today. It's the most useful logging change you can make this week.
3. **Your active user count** for the same month. Use people who actually did the core action, not everyone with an account. If you don't have a clean definition yet, this guide on [tracking product metrics before your first 100 users](https://indiehunt.io/studio/how-to-track-product-metrics-before-100-users) helps you pick one.

No per-user logs yet? Estimate from the total bill and your heaviest accounts' request counts, label the sheet as an estimate, and redo it next month with real data.

### Build the sheet

One row per active user: requests, input tokens, output tokens, model, AI cost, plan price. Sort by AI cost and group into buckets. The shape matters more than the average, because the long tail is where your margin goes.

![earlyhunt-2026-10-05-inline-1.jpg](https://txmhk1zrnc.ufs.sh/f/xSkWTCqmKWx9hwtwF6HRJkXTmPie5h04ySvxCrwdZbuRtDGB)

Here's what that looks like for a made-up tool. Call it DraftPilot: an AI assistant that writes cold-email drafts, sold at $19 a month on one flat plan. **Every number below is illustrative, not real data.**

Segment (illustrative)

Users

Avg AI cost per user / month

Segment total

Margin on AI cost at $19

Light

120

$0.40

$48

98%

Regular

64

$2.50

$160

87%

Heavy

12

$14.00

$168

26%

Top

4

$36.00

$144

-89%

**Total**

**200**

**$2.60 average**

**$520**

**86% blended**

The blended margin, 86%, looks healthy. Now look at the bottom rows. Four users, 2% of the base, eat 28% of the bill, and each one costs nearly twice what they pay. The 16 heavy and top users together are 8% of users and 60% of AI spend. The average hides all of that, which is why you sort the sheet before you trust it.

Last step for the sheet: work out the **cost per unit** your users actually understand. For DraftPilot that's one draft. Divide AI cost by drafts generated, including retries. Say it comes to about $0.01 per draft. That number is what you'll price around.

## Step 2: Pick a margin floor

A margin floor is the line you won't let AI cost cross on any plan or pack. It's a decision, not a benchmark. Remember AI isn't your only cost: hosting, payment fees, support time and refunds come out of the same dollar.

![earlyhunt-2026-10-05-inline-2.jpg](https://txmhk1zrnc.ufs.sh/f/xSkWTCqmKWx9ztMdLJN6rVYGydnu9LqWaXMOH7J3KS2gD65P)

DraftPilot's rule (again, illustrative): **AI cost should stay at or under 30% of what a customer pays.** That turns pricing into simple arithmetic:

- Max AI budget per plan = plan price × 30%
- Included usage = max AI budget ÷ cost per unit, rounded down

For the $19 plan: $19 × 0.30 = $5.70. At $0.01 a draft, that's 570 drafts. Round down to 500 included drafts a month, which leaves room for a bad month. The light and regular users from the sheet sit far below 500, so nothing changes for most customers.

Leave a buffer below the line. Cost per unit drifts as prompts get longer and features get added.

## Step 3: Choose your pricing shape

There are three practical shapes for an early AI product. The [Stripe usage-based billing docs](https://docs.stripe.com/billing/subscriptions/usage-based) are a useful reference for how each can be billed, including metered usage, prepaid credits and usage alerts.

Shape

How it works

Fits when

Watch out for

**Flat plan + fair-use cap**

One monthly price, a stated monthly limit

Usage is fairly even and buyers want a predictable bill

Vague "fair use" wording; caps that surprise people mid-task

**Credits or pure usage**

Buy a pack of credits, or pay per unit used

Usage is spiky or occasional; buyers are technical

Bill anxiety; people ration the product and never build a habit

**Hybrid**

Base plan with included usage, plus top-up packs or a higher tier

Most users are light but a tail is heavy (most AI tools)

Too many tiers and pack sizes; keep it to two plans and one pack

### Flat plan with a fair-use cap

Simplest to explain and easiest to buy. It works if your sheet shows usage fairly evenly spread. The cap stops the rare extreme case, not regular users. If more than a few percent of users hit it each month, the cap is too low.

### Credits or pay-as-you-go

Your margin is protected on every unit. The cost is psychological: when every click visibly spends money, people use the product less, and a tool nobody uses doesn't get renewed. Credits suit occasional jobs better than daily habits.

### Hybrid: base plan, included usage, top-ups

This is where most early AI tools end up, and it's where DraftPilot lands:

- **Starter, $19/month:** 500 drafts included
- **Pro, $49/month:** 1,400 drafts included (about $14 of AI cost, 29%)
- **Top-up:** 500 extra drafts for $17 on either plan (about $5 of AI cost, 29%)

Every piece holds the 30% floor. The heavy users from the sheet fit Pro. The four top users, at around 3,600 drafts each, would pay $49 plus five top-ups, $134 in total against roughly $36 of AI cost. They go from losing money to being your best accounts, and nobody got banned.

## Step 4: Word your limits honestly

Limits feel like a trap when they're hidden, vague or enforced by surprise. Write them so a buyer could predict their bill before paying.

- **Use a unit people understand.** "500 drafts a month," not "2 million tokens." Nobody outside your team knows what a token is worth.
- **Put the number on the pricing page,** not in the terms.
- **Tell people where typical usage sits,** but only if your sheet backs it up. "Most people use under 250" is useful context. Making it up is not.
- **Say what happens at the limit.** Does the tool stop, slow down, or offer a top-up? Pick one and say it.
- **Warn before the wall.** An email or in-app notice at 80% of the limit turns an angry support ticket into an easy upgrade.

DraftPilot's pricing-page line could read: *"Includes 500 drafts a month. Most customers use under 250. We'll email you at 400. If you run out, add 500 more for $17 or switch to Pro. Nothing renews or charges automatically."* Only promise "nothing charges automatically" if your billing really works that way.

## Step 5: Handle the 1–5% power users

Every AI tool gets a few accounts that use far more than everyone else. Don't treat them as a problem until you know who they are. Look at your top rows and sort them into three groups:

1. **Your best customers.** They've built the product into real work. Email them personally, ask what they use it for, and offer a bigger plan. These conversations often show you what your Pro tier should be.
2. **Shared or scripted accounts.** One login used by a whole team, or an automation hammering your API. Offer a team plan or an API plan, and add sensible rate limits.
3. **Abuse.** Free-tier farming, resold access, prompts unrelated to your product. Enforce your terms and move on.

For the first two groups you can also cut cost without touching price. Route simple tasks to a cheaper model, cache prompt parts that repeat, and move work that doesn't need an instant answer to batch processing, which both OpenAI and Anthropic list at lower rates. Check your provider's current numbers, then re-run the sheet.

## Step 6: Plan for model price changes

Model prices move. New models launch at different prices and older ones get retired. Build for that from the start.

- **Keep your unit abstract.** Sell drafts, minutes or reports, never "GPT-whatever calls." Then you can switch models without rewriting your pricing page.
- **Re-measure cost per unit after any model change,** not only when list prices change. A new model can use more or fewer tokens for the same job. Anthropic's pricing docs, for example, note that its newer models use a tokenizer that produces more tokens for the same text.
- **When costs drop,** choose on purpose: keep the margin, raise included usage (an easy, visible win for customers), or move to a better model at the same price. Raising included usage builds the most goodwill.
- **When costs rise,** change limits for new signups first. Give existing customers clear notice and a date. Never shrink a limit quietly. People notice, and they talk about it.

## Should you offer BYOK (bring your own key)?

BYOK means the customer plugs in their own API key and pays the model provider directly. Your inference cost for that user goes to zero, and you charge for the workflow, interface and everything around the model.

It fits when your buyers are developers who already have provider accounts, or when a few power users want more usage than you can afford to include. It's a poor fit for non-technical buyers: creating an API account, adding a card and pasting a key is a lot of setup before they see any value.

If you offer it:

- Sell it as its own plan with a lower platform fee, not a hidden setting.
- Store keys encrypted, never log them, and let users revoke them in one click.
- Expect support tickets that are really provider problems, like expired cards or rate limits. Write a short help page for them first.
- Say which models you support, because your prompts were written for specific ones.

## A 14-day pricing sprint

**Days 1–2: Pull the data.** Last month's provider bills, per-user request logs and active user count. Start logging token counts per user if you aren't already.

**Days 3–4: Build the sheet.** One row per active user, sorted by AI cost, grouped into buckets. Work out cost per unit, including retries.

**Day 5: Set the floor.** Pick your margin floor and calculate included usage for your current price.

**Days 6–7: Pick the shape and write the page.** Draft the pricing copy with the honesty rules above. Two plans, at most one top-up pack.

**Days 8–9: Add metering and warnings.** Show remaining usage in the app. Send the 80% email. Make sure what happens at the limit matches what the page says.

**Day 10: Talk to the top users.** Email your five heaviest accounts before anything changes for them. Ask how they use the product and offer them the right plan.

**Days 11–12: Ship to new signups.** New pricing applies to new customers first. Existing customers get notice and a date, or stay grandfathered for a set period.

**Days 13–14: Review.** Check margin per plan, limit hits, top-up purchases and complaints. Adjust one thing, then wait a full month before touching it again.

## Common traps

1. **"Unlimited" on an AI product.** Someone will take you up on it.
2. **Pricing in tokens.** Buyers can't judge it. Price in the outcome they care about.
3. **Trusting the average.** A healthy blended margin can hide a tail that loses money on every account. Sort the sheet.
4. **Forgetting hidden calls.** Retries, regenerations, embeddings, moderation checks and background summaries all cost money. Count them in cost per unit.
5. **A free tier on your most expensive model.** Free users with heavy usage can sink you before anyone pays. If you're torn on the free model itself, see [how to choose between a free trial and freemium before product-market fit](https://auraplusplus.com/studio/how-to-choose-free-trial-vs-freemium-before-pmf).
6. **Surprise bills.** Uncapped overage on a card nobody is watching turns a happy customer into a chargeback. Cap it or ask first.
7. **Quietly shrinking limits.** Cheaper for a month, expensive for your reputation.

## The short version

Measure cost per active user and sort it. Pick a margin floor. Sell usage in a unit buyers understand, include enough that most people never think about it, and give the heavy tail a fair way to pay more. Put the limits on the pricing page in plain words and review the numbers monthly.

You don't need perfect pricing to launch. You need pricing that won't lose money on your best customers. Once yours holds up, putting your AI tool in front of early adopters on [EarlyHunt](https://earlyhunt.com) is a good way to find out whether real users land where your sheet says they will.

## Links

- Page (HTML): https://earlyhunt.com/studio/how-to-price-an-ai-product-when-costs-vary-per-user
- AI-friendly Markdown: https://earlyhunt.com/studio/how-to-price-an-ai-product-when-costs-vary-per-user.md

## Explore

- [Studio](https://earlyhunt.com/studio)
- [Blog](https://earlyhunt.com/blog)
- [Launch Guide](https://earlyhunt.com/launch-guide)

_HTML version: https://earlyhunt.com/studio/how-to-price-an-ai-product-when-costs-vary-per-user_