AI generation costs: how do credit and token systems work?

Short answer: AI tools usually price generation in an internal unit such as tokens or credits instead of money, because every generation is real computation and its cost depends on the size of the job. A video costs more than an image for a simple reason: a video produces dozens of frames, an image produces one. This article explains how credit systems work, what drives the cost and how to spend less. It is brand-neutral; we do not give the prices of any specific tool.

What is a token or credit system, and why does it exist?

A token or credit is the internal currency you spend on generation inside a tool. A single unit makes it easy to price different jobs: a cheap image might cost a few units, a long video hundreds. For the tool provider, it also lets them adjust ratios when the real cost behind the scenes (processors, memory, model licences) changes, without rewriting every price tag.

You will usually see two routes: one-time packs (buy once, use until gone) and monthly credits included in a subscription. Because rules differ in every tool, read these three things before buying: whether credits expire, whether unused monthly credits roll over to the next month, and whether credits are refunded if a generation fails.

How do providers calculate the price?

Behind a tool's credit table there is usually the pricing model of the model providers. The official pricing pages show the outline of that model:

• Images, in tokens: According to Google's Gemini API pricing page, image generation models are priced per token, and an output image is converted into a token count based on its resolution. One note on the page says a 1K (1024x1024 pixel) output image consumes 1120 tokens. OpenAI's pricing page also prices image models per 1 million tokens and points to a calculator for per-image estimates.
• Video, in seconds: A footnote for a video model on the same Google page says billing is based on total output tokens, calculated at 5,792 tokens per second of 720p video. So video cost rises with seconds and resolution.
• Tiers: The same page lists separate prices for tiers such as standard, batch, flex and priority; accepting a wait goes cheaper, asking for fast responses costs more.

These figures change over time; check the relevant page for current values. What matters here is the logic: cost is proportional to the amount of computation.

Why does video cost more than an image?

Because a video is made of dozens of frames per second and the model generates them together. The Movie Gen paper shows the scale: the largest model's context is at most 73 thousand video tokens, which corresponds to roughly 16 seconds at 16 frames per second. A single image is 1120 tokens in Google's example. That ratio explains why video generation is priced many times above image generation. For a closer look at how video works, see our article How does AI video generation work?.

What drives the cost?

Factor	Effect on cost	What you can do
Model	Larger, higher-quality models usually cost more credits	Draft with a light model, finish with a strong one
Resolution	Higher resolution means more tokens	Try at low resolution first
Duration (video)	Every extra second adds cost	Do not ask for more than needed, join short clips
Audio	Generation with sound may be priced separately by provider	Turn it off if you do not need it
Number of attempts	Every regeneration is a new cost	Sharpen the request and go with few attempts
Input images	Reference images may count as input tokens	Add only what helps

8 habits to spend less

1. Start small: test the idea at low resolution or with a fast model and scale up only the one you like.
2. Write the prompt before you try: a vague request means many attempts. The rules in our How to write a good prompt article help here.
3. Ask for one change at a time: changing five things at once costs five times as much when it misses.
4. Start with an image, then move to video: animating a frame you like usually needs fewer attempts than trying a video from scratch.
5. Shorten the duration: a 5-second clip is enough for most posts.
6. Note the seed and settings: reuse a setup that worked.
7. Look at batch options: if you are not in a hurry, some providers' batch or flex tiers can be cheaper.
8. Track your balance: know whether monthly credits roll over so they do not go to waste at month end.

Example budget logic: first 3 fast drafts; if one is liked, generate only that one once at high resolution; if it fails, fix the request text instead of blindly repeating the same request.

What to check before you buy

• What the unit price is: how many credits for an image, a second of video and an edit?
• Expiry and rollover rules.
• Refunds for failed generations: what happens when a generation errors out or hits a filter?
• Subscription cancellation: how do you cancel, and can you use it until the period ends?
• Commercial use rights: paid credits do not always grant commercial rights; read the tool's terms.

Subscription or pack? A hypothetical calculation

Which is cheaper depends on your usage. The numbers below are invented units, only to show the logic; they are not the price of any tool. Imagine a tool where an image costs 4 credits and a five-second video costs 60. To test an idea you generate three draft images (12 credits) and one video of the one you like (60 credits), which is 72 credits. If you need to try the video twice, the spend rises to 132 credits; video attempts drain a budget much faster than image attempts.

So first think about how often you will generate. If a few images a month are enough, a small pack or a low monthly credit allowance does the job. If you make video regularly, work out how much of the monthly credit corresponds to seconds of video. If monthly credits do not roll over, avoid paying for credits you will not use. For packs, check whether there is an expiry date.

A simple rule: from the tool's price table, work out how many credits one image and one second of video cost, multiply by your own monthly output and compare with the total credits. Always compare against a real pack, not a guess.

Costs that are easy to miss

The sticker price does not tell the whole story. Failed or unwanted generations often also deduct credits, so the real unit cost is the price multiplied by the average number of attempts you need for a usable result. For example, even if a credit pack looks cheap, if the tool's results do not suit you, making twice as many attempts per job doubles the cost. When comparing two tools, look not only at the unit price but also at how often you get a usable result on the first try.

Also account for three things: whether later steps such as upscaling to higher resolution ask for separate credits, whether extra features such as editing and regeneration are priced separately, and whether commercial use needs a separate plan. The answers are usually written in the footnotes of the pricing page or in the terms of use.

Common mistakes

• Choosing the highest setting in the draft phase: You pay a lot for something you may not keep.
• Asking for longer video than needed: Every second costs, and most posts do not need it.
• Resending the same vague request: Fix the text first.
• Not thinking of credits as real money: Work out the unit price of the pack.
• Missing the expiry date: Unused credits can disappear.

Summary

• A credit system is an internal currency that prices different jobs with a single unit.
• Images convert to tokens by resolution and video by seconds and resolution; official pricing pages show this logic.
• Video is clearly more expensive than images because many frames are generated together.
• Model, resolution, duration, audio and number of attempts drive the cost.
• Start small, be specific, keep it short and read the rules before buying.