Deep-dive guide
AI Credit Pricing Simulator: Margin, Breakage and the Cost Behind Each Credit
Executive industry analysis
AI features have a cost structure that old software did not: every use consumes real money in model inference. A document summary, an image or a voice clip each costs a measurable amount, and heavy users can cost far more than light ones. Flat subscriptions hide this, which is why many AI products moved to credits. A credit turns a messy mix of tokens and compute into one number that customers can understand and that you can price with margin.
The structure of a credit system is a chain of conversions. Customers buy credits at a price per credit. Each task consumes a number of credits, so the price of a task is credits per task times price per credit. Behind it sits the cost of the task, which is set by models you do not control and which can change with a provider's price list. Margin per task is the gap between those two numbers, and everything else, from support to platform, is paid out of that gap.
The domino model follows the order. The number of credits redeemed, after breakage, determines tasks run. The margin per task decides gross margin once failed tasks and payment fees are counted. Overhead such as support, goodwill refunds and the platform eats a share of what is left, and the last box is profit. A model price increase or a customer shift towards heavier tasks hits the second box first, so margin can collapse while revenue and task counts look healthy.
The iceberg shows what a simple per-task margin misses. Visible are the inference cost and platform. Below the waterline sit retries and regenerations that burn inference with no extra revenue, payment fees, support for people who are confused about credits, and goodwill credits given to calm complaints. Breakage deserves its own caution: unused credits lift margin, but a model that depends on them can become a trust problem. Use the simulator to test credit prices and see where profit reaches zero.
Formula and methodology blueprint
The model is a single month. Credits sold are customers times credits per customer, and revenue is credits sold times price per credit. Redeemed credits are credits sold minus breakage. Tasks run are redeemed credits divided by credits per task. Inference cost is tasks times cost per task, plus a share for retries. Payment fees, support per customer, goodwill credits and a fixed platform cost complete the picture.
- Revenue
- R = customers × credits per customer × price per credit
- Cash value of the credits sold in the month.
- Tasks run
- T = credits sold × (1 − breakage) ÷ credits per task
- How many AI tasks customers actually execute.
- Price per task
- p = credits per task × price per credit
- What the customer pays each time a task runs.
- Inference cost
- I = T × cost per task × (1 + 8% retries)
- Model spend, including tasks that fail or are regenerated.
- Net profit
- Net = R − I − fees − support − goodwill − platform
- The monthly result. The breaking point is where it reaches zero.
Cascading elasticity
Cost per task and credits per task are the most elastic inputs, because they set the margin on every task at once. Price per credit scales revenue and margin together. Breakage helps profit but is the least reliable lever, since it depends on customer behaviour you cannot fully control.
Three real-world case studies
Bootstrapped starter
A new tool with a few hundred customers
A young AI tool sells small bundles at a healthy price per credit and uses a cheap model for most tasks. Margin per task is good, and the fixed platform cost is covered. Loading it shows the early stage working when credits are priced above the real cost.
Scaling mid-market
A scaled product with thousands of customers
A product with nine thousand customers runs large volumes of tasks at a lower price per credit but also lower cost per task. Volume carries the platform and support costs. Loading it shows how scale and routing together produce a large profit.
Black-swan crisis
A price cut meets a costly new model
A product cuts credit prices to win share while switching to a more expensive model and charging few credits per task. Customers use nearly every credit. Loading it shows every task losing money and which lever fixes the margin.
Industry benchmark matrix
| Metric | Danger zone | Industry average | Top 10% elite |
|---|---|---|---|
| Gross margin on AI tasks | Below 30% | 50–70% | Above 75% |
| Credit breakage rate | Above 40% (trust risk) | 10–25% | 5–15% with clear expiry rules |
| Model cost as share of task price | Above 60% | 25–45% | Below 20% |
| Retry and regeneration share | Above 15% | 5–10% | Below 4% |
| Support and platform as share of revenue | Above 45% | 20–30% | Below 15% |
Variables and levers
Paying customers
Customers who buy a credit bundle in the month.
Lever: Grow customers only after unit economics per task are positive, because volume at a loss just scales the loss.
Price per credit
What a customer pays for one credit on average, after any bundle discount.
Lever: Raise the price per credit or shrink bundle discounts; credits are an abstract unit, so small increases are easy to explain.
Credits sold per customer
Average credits each customer buys in the month, across plans and top-ups.
Lever: Encourage larger bundles with modest discounts and auto top-up, which raises revenue per customer for the same support cost.
Credits never used (breakage)
Share of purchased credits that expire or sit unused. Higher breakage raises margin but can hurt trust.
Lever: Treat breakage as a bonus, not a plan: clear expiry rules help margin, but heavy breakage invites refund demands and churn.
Model cost per task
What one AI task costs you in inference, including tokens, image or audio processing and third-party calls.
Lever: Route simple tasks to cheaper models, cache repeated results and cap output length to cut the cost of every task.
Credits charged per task
How many credits the customer spends each time they run one task. More credits per task means a higher price per task.
Lever: Price heavy tasks at more credits than light ones, so that the credit cost follows the real cost of the task.
How to use this simulator
- 1Start from the realistic preset or enter your price per credit, credits per task and cost per task.
- 2Read the five domino boxes to see whether margin per task or overhead turns red first.
- 3Press Find my breaking point to see the model cost or price that wipes out profit.
- 4Compare the current plan with the repricing plan on the 24-month chart.
- 5Review real usage monthly and reprice credits whenever model costs or task mix change.
Mistakes to avoid
- Pricing credits from average cost and ignoring that heavy users and heavy tasks cost far more than the average.
- Counting breakage as guaranteed profit, when it varies with behaviour and can trigger refunds and complaints.
- Forgetting retries and regenerated outputs, which add inference cost without adding revenue.
- Keeping credit prices fixed when the model provider changes its price list.
Frequently asked questions
Why use credits instead of a flat subscription?
Credits tie price to usage, so heavy users pay for the cost they create, and they give you a unit you can reprice when model costs change. Flat plans are simpler but can lose money on heavy users.
Is it fair to profit from unused credits?
Selling prepaid credits that some customers never use is common, but expiry and refund rules should be clear and compliant with consumer rules in your market. Check current rules for your region.
How do I set credits per task?
Start from the cost per task, add your target margin, then express that price in whole credits. Charge more credits for heavier tasks so price follows cost.
What if my model provider cuts prices?
You can keep prices and take the margin, or reduce credits per task to pass savings on. The simulator lets you test both.
How do I find the cost at which I lose money?
Press Find my breaking point. The tool moves each input towards its worse end and reports the value where monthly profit reaches zero.
How accurate is this?
It is an educational model with stated assumptions, not financial advice. Replace the inputs with your real usage and billing data.
This is an educational planning model with stated assumptions. It is not financial, tax or legal advice. Replace the inputs with your own figures and verify important decisions with a qualified professional.