How to Price AI Credits: Cost, Margin, Packs and Free Tiers

How to price AI credits

To price AI credits, work out what each action costs you in model fees plus overhead, decide what one credit is worth to you after store fees, and charge each action a whole number of credits that keeps your target margin at your cheapest price per credit. Then size the free tier as a marketing budget, set expiry rules per type of credit, and let free users earn more credits at a rate tied to what you earn. The rest of this guide walks through each step with illustrative numbers you can replace with your own.

If you are still choosing between subscriptions, ads and credits at all, start with How to monetize an AI app or chatbot.

Step 1: find your real cost per action

Users never see tokens. They see actions: a chat reply, an image, a minute of audio, a document summary. Price those actions, not tokens, and measure each one in your own logs.

For text, cost per action is input tokens times the input price plus output tokens times the output price. On Anthropic’s Claude API pricing page, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and Claude Sonnet 5.5 costs $2 and $10. The same page lists a 50% discount for the Batch API and says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text. For images, Google’s Gemini API pricing page lists gemini-3.1-flash-image at $0.067 per 1K output image on the paid tier. All prices checked on November 5, 2026. They change often, so read the live pages before you set prices.

Model fees are not your whole cost. Add a buffer for:

• Retries and regenerations that users trigger when they dislike an answer.
• Failed or blocked generations that you still pay for.
• Hidden tokens: system prompts, tool calls, chat history, safety checks.
• Storage, bandwidth and payment fees per action.

Illustrative cost table

Sample assumptions only: a chat reply uses 2,000 input and 500 output tokens, an image is one 1K image, and overhead adds 20% on top of model fees.

The spread is the point: in this example one image costs as much as about 150 fast replies. That is exactly why credits work better than a flat message count when your actions differ in cost.

Step 2: set what one credit is worth and your margin

Pick a simple retail anchor, such as 1,000 credits for $9.99, so one credit sells for about one cent. Then work out what you actually keep. If you sell through an app store, take off the store fee; if you sell on the web, take off card fees and tax handling. In this illustrative case, a 15% store fee leaves about $0.0085 per credit.

Next, choose a target gross margin for inference: the share of each credit’s net value left after model and overhead costs. Many teams aim high because support, servers, salaries and refunds come out of what is left. The formula is:

credits_per_action = ceil( cost_per_action / (net_value_per_credit x (1 - target_margin)) )

Illustrative credit prices at a 70% target margin

At $0.0085 net per credit and a 70% target, each credit can carry at most $0.00255 of cost. Rounded to friendly numbers:

The image lands slightly under 70% because 30 credits reads better than 32. That trade is fine as long as you choose it on purpose. The fast reply has a large margin because 1 credit is the smallest unit; that is your room to be generous with the free tier.

Credit packs vs subscriptions

Most AI apps end up with both: a subscription that includes a monthly credit allowance, and packs for people who need more or do not want a plan.

For a subscription, price the allowance with the same formula, but assume users spend all of it. If a $9.99 plan includes 1,200 credits and every credit is spent on images, your cost in this example is 40 images x $0.0804, about $3.22, against about $8.49 net. That still works. A plan that only works if users leave credits unspent is fragile, because your heaviest users are the ones who renew.

Show the balance on every screen that spends credits, and show the cost before the action runs (“This image uses 30 credits”). Hidden credit math leads to surprised users and refund requests. Hybrid monetization covers how to keep plans, packs and free credits from competing with each other.

How many free credits to give

Treat free credits as a marketing budget with a cap. Decide how much you will spend to get one user to their first good result, then work backward.

• Size the welcome grant for one win. Enough for a user to finish one real task, such as 3 images or 30 fast replies, not enough to finish a project.
• Prefer a daily refill to a big monthly grant. In this example, 20 fast-reply credits a day costs at most about $0.011 per user per day, and it brings people back daily.
• Route free credits to cheap actions. Let free credits buy fast replies but not the expensive model, or charge more credits for costly actions on the free plan.
• Require an account and rate limits, so one person cannot claim the grant ten times.
• Set a spend alarm on daily inference cost for free users.

Then measure: what share of free users reach the cap, and what share of those upgrade or buy a pack? If almost nobody hits the cap, it is too generous. If most hit it on day one and leave, it is too tight. How to monetize free users covers what to offer people who will never pay.

Expiry rules: what can and cannot expire

Expiry is where credit pricing meets platform rules and consumer trust. Keep separate buckets for each kind of credit, because the rules differ.

• Purchased on iOS. Section 3.1.1 of Apple’s App Review Guidelines says credits or in-game currencies purchased via in-app purchase may not expire, and that apps must use in-app purchase to unlock features such as in-game currencies (checked on November 5, 2026). Price iOS packs assuming those credits live forever.
• Purchased on the web. Your payment platform may let you set expiry. Stripe’s billing credits docs say credit grants only expire if you set an expiry time, that promotional credits usually have an expiry date, and that billing credits cannot be offered as gift cards or stored value (checked on November 5, 2026). Local consumer law can also limit expiry on prepaid value, so check it for your markets.
• Subscription allowance. A monthly allowance that resets is common. Say clearly whether unused credits roll over.
• Free daily or welcome credits. These can reset, as long as you say so up front.
• Earned credits. Decide one rule and write it in your terms. Users worked for these, so harsh expiry feels unfair.

Spend the bucket that expires first, and show that order in the balance view. A clear order also makes refunds and reversals easier to handle. Offerwall rules for iOS apps covers the Apple side in more detail.

Letting free users earn credits through offers

Some free users will never pay but would complete an offer, such as trying an app or answering a survey, in exchange for credits. An offerwall pays you per completed offer, and you turn that payout into credits. The pricing question is simple: how many credits do you give per $1 you earn?

Use two checks:

1. Cost check. Find your most expensive cost per credit across all actions. In the illustrative table that is the image: $0.0804 / 30, about $0.0027 per credit. At 100 credits per $1 earned, the worst case is $0.27 of inference for each $1 you earn, leaving room for reversals and support.
2. Price check. Compare with what buyers get. At 1,000 credits for $9.99, buyers get about 100 credits per $1. Earning at roughly the buyer rate keeps packs worth buying while still being fair to people who gave their time.

With those numbers, an offer that pays you $0.80 would show 80 credits: about 80 fast replies or between 2 and 3 images. Sample figures only. The virtual currency calculator lets you test your own rate and split, and How to set your virtual currency exchange rate explains the formula and rounding. For the wider design of earn and spend loops, see Virtual economy design.

Handle reversals before you launch

Advertisers sometimes reverse a completed offer, and the credits you granted can already be spent on inference you paid for. Credit only from a server-side postback, take credits back on a reversal, and consider a short hold before earned credits can buy expensive actions. Reward hold periods and Publisher earnings and reversals cover both sides.

How to add Moonetix to your AI app

Moonetix fits AI web apps and mobile apps where users sign in and hold a credit balance on your server.

1. Apply as a publisher at moonetix.com. Every application is reviewed by hand; once you are approved, each app you add is live right away.
2. Add your app and set the currency: name it “credits”, add an icon, and set the conversion rate, user split and whole-number rounding from your cost and price checks above. Set up your app currency walks through the settings.
3. Integrate. For a web app, build a signed offerwall link or iframe on your server for the signed-in user (see the web apps page and website integration); for mobile, use the Android or iOS SDK with session tokens your server requests.
4. Credit only from the signed postback: status 1 adds credits, status 2 with the same transaction ID removes them. Test with “Open as a test user” and “Send a test postback” in the app’s Integration tab, following Test your app before launch.

Honest limits: offers depend on each user’s country and device, so some users will see few offers. Users on VPNs, proxies, Tor or data center networks are blocked from the wall, and end users must be at least 16 (with parental permission under 18), so an AI app for children is not a fit. Earned credits should add to your paid plans, not replace them.

Frequently asked questions