AI Video Tools: The Credit Maths Nobody Puts on the Pricing Page

The pricing pages for AI video tools are the least honest in software. They show a monthly number and a credit allowance, and they do not show what a credit buys, because if they did the comparison would be over in thirty seconds.

This is not a scam. It is a category where the underlying cost — GPU seconds — is genuinely hard to express as a per-feature price, so vendors picked the vaguest unit available. The result is that two tools with identical-looking plans can differ by a factor of five in what you actually pay per finished minute.

Here is how to work out the real number.

What a credit actually buys

Ask four questions of any credit system, in this order:

1. Is a credit a generation, a second, or a resolution? Some tools charge one credit per generation regardless of length, which is cheap for short clips and ruinous for anything over ten seconds. Others charge per second of output, which is predictable. Others charge per second *at a resolution*, so the same clip costs four times as much at 1080p as at 480p.

2. Do failed generations consume credits? Most do. If your hit rate is one usable clip in four — which is typical for anything with a specific subject — your effective cost is four times the sticker.

3. Does re-rolling a prompt cost the same as the first attempt? Usually yes, and this is where the money goes, because you will re-roll.

4. Do credits expire? Some roll over monthly, some expire at the end of the billing period, and some expire a fixed number of days after purchase. An expiring credit pack is a subscription with extra steps, and buying a large one on an annual plan can mean paying for capacity you will never use.

Write the answers down for each tool. The differences are large enough that the ranking usually flips.

Work backwards from a finished minute

The number you care about is not "cost per credit". It is cost per finished minute of usable video. Derive it like this:

1. Estimate your hit rate honestly. For a specific subject or a consistent character, assume one usable output in three. For generic b-roll, one in two.
2. Estimate the number of generations per finished second. A ten-second clip with a specific subject typically needs eight to fifteen generations.
3. Multiply by the per-generation cost at the resolution you actually need.
4. Divide by the length.

A worked example: a tool charging one credit per five-second generation at 720p, with a plan giving 1,000 credits for $30, and a hit rate of one in three. That is 30 cents per generation, 90 cents per usable five-second clip, and about $10.80 per finished minute at 720p. The same tool's competitor charging per second might come out at $3 a minute. The plan that looked cheaper is three times the price.

This arithmetic is unglamorous and it is the whole decision. Do it once, in a spreadsheet, for the three tools on your shortlist.

The avatar-video meter is different again

Avatar tools — the ones that put a presenter on screen — meter on a different axis, and the trap is not credits but video length per generation. A plan may offer unlimited videos and cap each one at ninety seconds. That is fine for social clips and useless for a twenty-minute training module, which then has to be assembled from fourteen pieces.

Check three things on avatar plans:

• Maximum length per generation. Not the total monthly minutes, the ceiling on one file.
• Whether translation counts against the same allowance. Usually yes, and translation is the feature people use most.
• Whether custom avatars are a separate tier. Often they are, and a custom avatar is usually the reason you chose the tool.

The watermark is a pricing feature, not a limitation

Free tiers watermark every export. That is not an accident and it is not primarily about preventing commercial use — it is a conversion mechanism, because a watermarked video cannot be shown to a client or published.

Which means the free tier is not a trial of the product. It is a trial of the interface. You cannot evaluate output quality from a watermarked export in any meaningful way, because the watermark occupies the attention of everyone watching it.

If you are seriously evaluating, pay for one month. It is the only way to see what the tool actually produces, and one month is cheaper than the week you will lose deciding.

The queue is a hidden cost

Credit-based tools run queues, and queue length varies by time of day and by plan tier. A tool that is cheap per minute but takes forty minutes per generation at 7pm has a real cost in your schedule, and it does not appear on any pricing page.

Test the queue at the time of day you will actually use it, on the cheapest plan you are considering. If a generation takes longer than the video you are making, the tool is unusable regardless of price.

Where this leaves the decision

Rank on cost per finished minute at your required resolution, then on queue time at your working hours, then on maximum generation length. Feature lists are close to identical across this category; the meters are not, and the meters are what you pay.

For the per-tool numbers, each AI video tool review records what the plan meters in and where the free tier stops being usable — including the ones whose credit cost makes them more expensive than they look. And if you are choosing between two specific tools, the HeyGen vs Synthesia and InVideo vs Pictory comparisons work through the arithmetic for those pairs.