Usage-Based vs Subscription Pricing for AI Products
Subscription is predictable; usage-based matches cost to value but is hard to forecast. How to choose, hybrid models, and the AI-specific cost problem.
Published 2026-09-17 · 4 min read
AI products face a pricing tension most software does not: the cost of serving a user varies enormously between users. One person runs ten queries a month, another runs ten thousand. A flat subscription charges them the same and loses money on one of them. Usage-based pricing matches cost to value but is harder to sell and harder to forecast.
Neither is universally right. This is how to choose, and the hybrids that get most of both.
The core trade-off
| Subscription | Usage-based | |
|---|---|---|
| Revenue predictability | High — you know next month | Low — it varies with customer behaviour |
| Cost alignment | Poor — heavy users lose you money | Strong — revenue tracks cost |
| Buyer preference | High — predictable budget | Mixed — hard to forecast a bill |
| Sales friction | Low — one number | Higher — requires explaining the meter |
| Expansion revenue | Requires tier changes | Automatic as usage grows |
| Churn signal | Explicit cancellation | Usage decline, which you can see coming |
When subscription is right
- Your cost per user is fairly uniform. If most users consume a similar amount, a flat price does not expose you to the heavy-user problem.
- The value is access, not volume. A tool someone logs into daily delivers value by being available, not by how many operations they run.
- You sell to people who budget. B2B buyers strongly prefer a predictable annual number they can put in a spreadsheet. Usage-based pricing creates an internal approval problem for them.
- You are early. A single flat price is simpler to sell, simpler to explain and simpler to change than a metered model you will get wrong.
When usage-based is right
- Cost per user varies by an order of magnitude. The defining case. If one user costs twenty times another, a flat price is a subsidy from the light users to the heavy ones.
- The value scales with volume. Processing more documents, generating more images, translating more minutes — the customer gets proportionally more value, so proportionally more charge is fair and defensible.
- You have an API or programmatic product. Developers expect metered pricing and are comfortable forecasting it.
- Your gross margin matters at scale. Usage-based protects margin as you grow; a flat price with variable cost gets worse with every heavy user you acquire.
The hybrids, which are usually the answer
| Hybrid | How it works | Suits |
|---|---|---|
| Subscription with an included allowance | Flat price includes N units; overage is metered | Most AI products — predictable base, protected margin |
| Tiered subscription by capacity | Higher tiers include more usage | Clear good/better/best where the buyer self-selects |
| Credit packs | Buy a block of credits, draw down with use | Variable or occasional usage; removes the subscription commitment fear |
| Subscription plus metered overage | Base fee for access, usage on top | B2B where access has value independent of volume |
| Per-seat plus per-usage | Seat fee for access, usage for volume | Team products where both headcount and activity matter |
Designing the meter
- Meter the thing the user understands, not the thing that costs you money. "Generations per month" is legible; "tokens" is not, even though tokens are your actual cost. Translate one into the other in a way the buyer can reason about.
- Pick a unit that maps to value. If the user gets more value per unit, charging per unit is defensible. If the unit is arbitrary relative to value, the meter will feel extractive.
- Make the allowance generous enough to reach value. A user who hits the limit before experiencing the product's worth will churn, not upgrade.
- Show usage continuously. A surprise bill is the fastest way to lose trust in a metered model. A visible counter and a warning before the limit converts a churn event into an upgrade conversation.
- Decide hard limit versus soft. Cutting off mid-task is hostile; allowing overage with a warning is friendlier and usually more profitable. Pick deliberately.
How to choose, concretely
- 1
Measure your cost spread
What does your heaviest user cost versus your lightest? If it is under about 3x, subscription is fine. If it is 10x or more, you need a meter or a cap.
- 2
Ask what the buyer budgets against
B2B wants a number. A developer with an API expects a meter. A consumer wants no surprises. The buyer's expectation constrains you more than your cost structure does.
- 3
Default to subscription with an allowance
Unless the cost spread is extreme, a flat price with a generous included allowance and metered overage captures most of the predictability and most of the margin protection.
- 4
Make the limit a cost ceiling, not a revenue cap
The allowance exists to stop you losing money on heavy users, not to maximise overage. Set it where value is demonstrated, not where extraction peaks.
What not to do
- Do not meter without showing the meter. A usage-based price the customer cannot see accruing is a bill they will dispute and a trust problem you will lose.
- Do not use a unit the customer cannot map to their work. Pricing per token to a marketing team is pricing in a language they do not speak.
- Do not set the allowance to maximise overage. It optimises a number while training customers to resent the product. Set it for value, then measure.
- Do not offer unlimited anything on a variable-cost product. "Unlimited" is a promise you are paying for on someone else's behalf, and the users who take you up on it are the ones who cost the most.
Subscription sells predictability; usage-based sells fairness. AI products usually need to sell both, which is why the allowance-plus-overage hybrid is where most of them land.
More in Pricing & conversion
- How to Launch a Paid AI Product With No Existing AudienceLaunching straight to paid with no list and no followers: why it can work, the trust you must build, and the sequence to a first paying customer.
- How to Run a Lifetime Deal Without Destroying Your PricingLifetime deals bring cash and users but can cap your revenue for good. When they make sense, how to structure one safely, and the traps you cannot undo.
- How to Price an AI Product: The Four Models and When Each FitsFree, freemium, paid and free-trial pricing for AI products: the cost structure each assumes, who each suits, and how to choose before you have users.