Pricing AI Video Plans Against Provider Costs

AI video generation cost per minute is the number your whole pricing rests on. Every video your customers make consumes minutes on a provider account you pay for, and that bill grows with use while your subscription price stays fixed. Set each plan's allowance from your measured cost per output minute, add the other costs on top, and make the server refuse jobs beyond the allowance.

This is an operator worksheet. It builds the unit economics step by step with labelled invented numbers, because we do not state any provider's price. Get a current quote from each provider you use, put the date on it, and rerun the tables. If you run a white-label HeyGen clone, the plan limits described here map onto settings the platform checks before it queues a job. For how the pipeline spends those minutes, read how an AI avatar video generator works.

Know your cost per output minute

Provider pricing pages and invoices use different units. Some bill by seconds or minutes of generated video, some by credits, some by a plan with included minutes and overage. Convert everything to one unit: cost per minute of finished output video, the minute your customer actually receives. Nothing else compares across providers.

How to measure it

1. Pick a test set. Twenty or more jobs that look like real customer work: different script lengths, avatars, voices and languages, with a share of premium options.
2. Run them through each provider you are considering. Use production accounts, not free demos, since demo behavior and limits can differ.
3. Add up the invoice. Take the total provider charge for the test period, including jobs that failed and jobs you re-ran.
4. Divide by finished output minutes. Do not divide by input script length or by minutes attempted. Divide by minutes that reached a customer.
5. Repeat per tier. Standard avatar, premium avatar, voice cloning and each translated language each get their own number.
6. Date it. Write the date and the provider plan next to the number. Prices in this field move.

Our platform records each generation attempt with the provider and the settings used, which makes the reconciliation easier: you can match provider charges to attempts instead of guessing. The HeyGen clone platform lets you change providers from admin settings, so you can run the same test set through two providers and compare, then keep the one that gives acceptable quality at a cost you can sell.

Invented starting numbers

The premium option costs about 2.4 times the standard one in this example. That ratio becomes the credit weight later.

Cost lines beyond the provider

The provider invoice is the biggest recurring cost but not the only one. Add each line to get a loaded cost.

Loaded cost per standard credit in the invented example: provider 0.70 times 1.08 for retries is 0.756. Add 0.03 storage and 0.02 delivery, and the loaded cost is 0.806, which we round to 0.81. Payment fees and support are per plan, not per job, so they come in the plan table. For the lines every operator forgets, read hidden running costs of a creator platform.

Storage deserves a decision before launch. A customer who cancels but leaves a large archive still costs you storage, so decide how long files are kept after a subscription ends, and write the rule into the plan terms. Our HeyGen clone development cost page covers storage modes and retention work as running costs you carry.

Setting plan allowances

Do not sell "minutes" if your jobs have different costs. Sell credits with weights, so every job type draws from one allowance and the worst-case cost per credit stays flat.

The translated row shows why weights need a check: 0.75 times 1.08 plus 0.05 storage and delivery is 0.86 per minute, so a translated minute at one credit costs more than a standard minute. Either weight it at 1.1 credits or accept the small difference. Use the highest loaded cost per credit, here 0.86, as your worst case for plan math, which is conservative.

Now set the plans. Margin is price minus payment fee minus support minus allowance times worst-case cost.

Check one row by hand. Pro: 99.00 minus 34.40 minus 3.27 minus 4.00 gives 57.33. That is the margin when the customer spends every credit. Because the allowance is a hard ceiling, the margin cannot fall below that figure no matter what the customer does. Average customers use less, and unused allowance is extra margin.

The invented margins here are high because the example ignores fixed overhead, sales cost and your own time. Subtract hosting, monitoring and staff before deciding the price is comfortable. A plan is safe when margin stays positive after overhead at your expected customer count.

Price against value, not only cost

Cost sets the floor. The ceiling is what the customer would pay to get the same video another way: a freelancer, a studio, or a hosted competitor. Look at what your target buyers pay today, then place your plan between your floor and their alternative. For who the buyers are, read who buys a white-label AI video platform. For the revenue models these plans fit into, see the HeyGen clone business model.

Limits enforced before the job

An allowance only protects you if the server enforces it. If limits are checked in the browser, anyone can bypass them with a modified request. If they are checked after a job runs, you have already paid the provider.

In our platform, generation limits are checked on the server in the service layer before a job is queued, and export entitlements gate resolution and format by plan, so overage is prevented rather than invoiced afterward. API keys are scoped to a workspace and usage is recorded per endpoint per day, which supports a metered developer tier. Queue behaviour matters too: job-queue libraries such as Bull offer concurrency control, priorities, retries and rate limiting, which are the building blocks for the concurrency and re-run caps above.

Two controls beyond the base limits are worth planning. First, a warning at 80% of the allowance, by notification, so customers upgrade before they hit the wall. Second, a platform-wide spend alert on your own provider accounts, so a bug or an abuse pattern that slips through shows up in hours, not at the end of the month. We can set up per-customer budget dashboards beyond what the admin console shows.

Stress test: what if provider prices move

Provider prices and models change often, so test your plans against a rise before it happens. Take the Pro plan from the table, with a worst-case provider cost of 34.40 at full use, and raise the provider cost line.

Even a doubling leaves this plan profitable before overhead, which is the sign of a plan with real headroom. If your own table goes negative at a 50% rise, cut the allowance or raise the price now. Keep a rule for what you do when cost moves: reprice new customers at once, give existing customers notice, and check your terms for what notice you must give. A second provider you have already tested is your best protection, because you can move traffic without waiting for a developer.

Annual plans and refunds

An annual plan collects cash early and cuts payment fees, but it also commits you to twelve months of provider exposure. Say Pro is sold for ten months' price, 990 up front. Release the 40 credits month by month, not all at once, so a customer cannot burn the year's allowance in a week and then ask for a refund. Payment fees fall from twelve charges of 3.27, which is 39.24, to one charge of 30.00, a saving of 9.24 in this example. Margin at full use is 990 minus 412.80 of provider cost minus 30.00 of fees minus 48.00 of support, or 499.20. Write the refund rule before you sell the plan: what happens to the unused months, and whether credits already used are refundable.

Worksheet

Fill this in with your own numbers. Do not skip a step.

1. Provider quote per tier. For each provider, write the current cost per finished output minute for standard avatar, premium avatar, cloning and a translated minute. Date each.
2. Retry rate. Failed or re-run jobs divided by total jobs, from your test set. Multiply provider cost by one plus that rate.
3. Storage and delivery per output minute. Average file size times your storage and bandwidth rates, spread over the retention period.
4. Loaded cost per credit. Provider cost times the retry factor, plus storage and delivery, divided by credits per minute. Take the maximum across job types.
5. Payment fee per plan. Percentage times price plus any flat fee, from your processor's terms.
6. Support allowance per plan. Expected support hours times hourly cost.
7. Plan allowance. Choose credits per plan, then compute worst-case margin as price minus fee minus support minus credits times loaded cost.
8. Overhead check. Subtract fixed monthly costs divided by expected customers. Margin must stay positive.
9. Overage price. Set a price per extra credit at least as high as the plan's effective price per credit, plus margin.
10. Limits. Set concurrency, maximum length, entitlements and API quotas to match the plan.
11. Review date. Put a monthly reminder to re-measure cost and reprice.

Overage and the developer tier

When a customer uses up the allowance, they either wait for renewal, upgrade or buy extra credits. Price extra credits above plan credits, or customers will skip the upgrade. In the invented example, Pro's effective price per credit is 99.00 divided by 40, or 2.475. An overage credit at 3.00, less a percentage payment fee of 0.09 (ignoring any flat fee) and a worst-case cost of 0.86, leaves roughly 2.05 of margin per credit, or 68%.

Metered API access is the same logic with a meter. Stripe describes usage-based billing and billing credits for prepaid or promotional usage, which are the standard tools for charging by use and granting prepaid or free credits. Note that Stripe's billing credits are for spending on your own products and services, not for stored value or third-party spending, so check what each tool allows before you design around it. Our platform models plans and a Stripe-compatible flow, and moving real money requires your own Stripe account and webhook wiring.

Free tier and trial design

A free tier is a marketing cost, so price it like one. Compute it and cap it.

If recovery takes longer than your typical customer stays, the trial is too generous or the conversion too low. Fix it by cutting trial credits, raising the quality of sign-ups or improving onboarding, in that order.

Protect the trial from abuse. Require email verification before the first job, optionally two-factor, and give one trial per workspace and per payment identity. Block voice cloning and premium avatars on trial, since they cost most and carry the highest consent risk. Cap the trial's video length. A scripted sign-up attack against an open trial could cost thousands in one night; a hard cap and verification make it a small, known number. Our platform supports email verification and optional two-factor at sign-up, and limits are checked server-side.

Consent and policy checks also belong on any plan that allows cloning or real-person avatars; see AI avatar consent, likeness and misuse.

What the platform price covers

Keep two budgets apart in your head. The platform price is the published, one-time price for the software, full source code, 60 days of technical support and a year of updates. It is not a running cost. The running costs are provider usage, storage, delivery, payment fees, hosting, monitoring and support, and they recur and scale with customers.

• One-time: the software package, rebranding and deployment. See our pricing for the published figure.
• Recurring and variable: provider minutes, storage, delivery, payment processing.
• Recurring and fixed: hosting, database, Redis, monitoring and staff time.
• Yours to arrange: provider accounts and approvals. We connect your providers; we do not supply render minutes or a bundled avatar and voice catalog.

We take no share of what your customers pay you. That means your margin per plan is the one you compute here, with no revenue share to subtract. It also means the cost discipline is yours.

What to do next

Run the measurement first. Pull twenty representative jobs through your candidate providers, compute the cost per finished output minute and date it. Then build the plan table, test the worst case and fix the limits. Only after the numbers hold should you publish prices.

When the table works, review the HeyGen clone features to confirm each limit you need exists, and keep a monthly review on the calendar, since models and prices change quickly. This article is general information, not financial or legal advice; take your terms of service, refund rules and tax treatment to a qualified adviser before you charge customers.