# Price testing for AI apps: find the price per credit that earns the most

Source: https://leverprice.com/for/ai-apps
Updated: 2026-11-17

> Test the pack price with the credits fixed, and optimize for profit. Leverprice shows each visitor one price from your range and follows them to a paid order. With the profit goal it subtracts provider fees and the unit cost you set, so a price that ties on revenue but keeps more after model costs wins.

An AI app pays for every request its users make. A heavy user on a cheap plan can cost more in model calls than they pay you, and the price that brings in the most revenue can leave the least after those calls.

Credits add a second layer of guessing. Buyers can't compare 1,000 credits with anything on a competitor's page, so the price per credit tends to come from your own API bill plus a markup that felt right. [AI app pricing with credits](https://leverprice.com/blog/ai-app-pricing-credits) compares packs, plans and hybrids; a price test settles the numbers inside the model you pick.

## Why is pricing an AI app hard?

Because every sale carries a real unit cost, so revenue and profit can disagree about which price is best. Example: a 1,000-credit pack. An average buyer's 1,000 credits cost you $4.00 in model calls, and your provider takes 5% + 50¢ per sale (Polar's starter rate as of November 2026, per its [fee page](https://polar.sh/docs/merchant-of-record/fees)).

| Pack price | Conversion | Revenue per visitor | Profit per sale | Profit per visitor |
|---|---|---|---|---|
| $10 | 3.0% | $0.300 | $5.00 | $0.150 |
| $15 | 2.0% | $0.300 | $9.75 | $0.195 |

Profit per sale = price × (1 - 0.05) - $0.50 - $4.00, which gives $5.00 at $10 and $9.75 at $15. Both prices earn the same revenue per visitor, and the $15 pack earns 30% more profit per visitor.

Leverprice's profit goal measures exactly this: revenue minus provider fees and the unit cost you set on the product. Under the revenue goal, these two prices would look like a tie. The [payment fee calculator](https://leverprice.com/tools/payment-fee-calculator) gives you the fee side for your provider.

Set the unit cost to what an average buyer actually burns. Example: if a fully used 1,000-credit pack costs $5.70 in model calls and buyers use about 70% of their credits, the expected cost is about $4.00.

## What should an AI app price test first?

The pack or plan most of your buyers start with. Keep the number of credits fixed and test only the price, so the result means one thing.

### 1. The starter credit pack

Example: 500 credits at $5, $7 or $9, which works out to 1.0¢, 1.4¢ or 1.8¢ per credit. Use the profit goal with the expected model cost of 500 credits as the unit cost.

### 2. The top-up pack for existing users

People who already use the product may react to price differently from first-time visitors, so test top-ups separately. Example: 2,000 credits at $15, $19 or $25. Pass the user id to `getPrice()` so each user sees the same top-up price on every device.

### 3. The monthly plan with included credits

Example: 1,000 credits a month at $20, $25 or $30. A plan is a subscription, so each price is judged on its first payment and renewals don't count toward the test.

## How does Leverprice fit an AI app's stack?

On your server, next to the code that creates the checkout. The TypeScript SDK runs on Vercel and Netlify functions, Cloudflare Workers, AWS Lambda, Bun, Deno and Node 20+. The Python SDK has sync and async calls for FastAPI, Django and Flask ([FastAPI guide](https://leverprice.com/guides/fastapi)), and there are SDKs for PHP, Ruby and Go. No Leverprice JavaScript runs in the browser.

Credit packs are one-time payments, which Stripe Checkout, Polar, Lemon Squeezy and Dodo Payments all confirm through signed webhooks. You paste a signing secret, never an API key, and Leverprice never sees card data. Refunds come off the revenue of the price that sold the pack, so refunded credits show up in the result. Lemon Squeezy closed to new stores on November 16, 2026 and stops processing payments on March 31, 2027 ([announcement](https://www.lemonsqueezy.com/blog/saying-goodbye-to-lemon-squeezy)).

[Regional prices](https://leverprice.com/blog/regional-pricing) set a price per country the way App Store Connect does: countries fall into 5 tiers by World Bank price level, and each tier pays a share of the standard price, converted to the local currency and rounded. Your model costs don't shrink with the local price level, so check the bottom tier against your unit cost. Example: under the balanced preset (100/85/70/55/40%), tier 5 pays the equivalent of $4.00 for a $10 pack whose credits cost you $4.00 in model calls. Set those countries by hand or pick a gentler preset; the `regional` guard's minimum for each currency makes sure nothing below it is ever served.

If you build with a coding agent, the dashboard's Install page has a ready prompt for Claude Code, Codex or Cursor, and `npx leverprice setup` connects your agents to Leverprice's MCP server. The agent can then create products, start and stop tests, read results and lock in winners, acting as you, with every change in the audit log; a read-only token limits it to reading.

## How do you set up a credit pack price test?

1. Create a project and add the pack as a product with its currency, fallback price and range, for example `fallbackPrice: 1000`, `minPrice: 800` and `maxPrice: 1800`.
2. Set the product's unit cost and provider fee through the API or your AI agent over MCP, and choose the profit goal.
3. Call `getPrice("credits_1000", { request, userId })` where the pack is shown and again at checkout. Charge `quote.price` in `quote.currency`, pass `quote.metadata` whole and append every `quote.setCookie` value; the [docs](https://leverprice.com/docs) have the code for each language.
4. In Payments, connect your provider's webhook signing secret.
5. Buy a pack in test mode to check the purchase arrives, then switch to the live key and start.

## Which Leverprice plan fits an AI app?

| Plan | Price | What you get |
|---|---|---|
| Free | $0 | Manual A/B split, 1 product, 1 experiment, up to $500/month in tested revenue |
| Pro | $29/month, or $199/year (about $17/month) | Automatic optimization, unlimited tested revenue, up to 20 sites |
| Lifetime | $299 once, 50 spots | Everything in Pro, forever |

Free is enough to test one pack by hand. Move to Pro when pack sales pass $500 a month or you want the engine to shift traffic for you: on $3,000 a month of pack sales, a 1% lift covers Pro's $29. Paid periods are non-refundable, so start on Free or monthly.

## Frequently asked questions

### Should I test the price per credit or the pack price?

The pack price. Keep the credit count fixed and the price per credit follows. If you change both at once, you can't tell which one moved buyers.

### Can I test usage-based pricing?

Not directly. Leverprice tests the price charged at checkout, so prepaid packs and plans work, while a per-token rate billed on next month's invoice never passes through a checkout. Prepaid credits are the testable form of [usage-based pricing](https://leverprice.com/glossary/usage-based-pricing).

### Can I test subscription plans as well as packs?

Yes. Polar verifies a subscription's first payment through its signed webhook. Stripe verification covers one-time Checkout Sessions only, so with Stripe subscriptions you report the first payment with `trackPurchase()` and it shows as unverified. Set the project's Learn from setting to Reported so the engine learns from it. Dodo Payments verifies one-time payments only.

### Does Leverprice see my users' prompts or usage?

No. Your server asks for a price and can pass a user id, and only a hash of it leaves your server. Purchases come from your provider's signed webhook, and Leverprice never sees card data or provider API keys.
