Before anything else
AI vendor pricing changes often, and any number written here will age. This piece teaches you how to do the arithmetic and shows the orders of magnitude between the three kinds of generation. Before deciding anything based on it, open your vendor pricing page and substitute the values.
The month we are budgeting
Let us use a realistic month for a small business that takes its own presence seriously: 20 pieces, distributed like this:
- 12 single-image posts, with captions
- 4 carousels of 5 cards, with captions
- 4 reels of 8 seconds, with captions and scripts
That is roughly one piece per working day. Plenty of people publish less; almost nobody sustains more without a machine behind them.
Line 1 — text is noise in the budget
Text is the cheapest generation there is, by a margin that usually surprises anyone who has never done the arithmetic.
A finished caption is, say, 120 words. The prompt that generates it is not just the instruction: it carries the brand kit — personas, tone of voice, banned terms, the offer with price and link, the compliance disclaimer. A well-assembled prompt like that lands between 800 and 2,000 tokens. The answer, around 300.
Twenty pieces, with reel scripts and variations, come to somewhere between 40 and 60 text calls in a month. Multiply that by any mid-cost text model of the current generation and you land in the range of cents to a few dollars for the entire month — and the cheaper models drop that by an order of magnitude.
Practical conclusion: it is not worth optimising the text line. If somebody offers you savings by changing the text model, the savings are real and irrelevant. Use the model that writes best in your language.
Line 2 — images are the middle ground
Model-generated images cost per piece, not per token, and the price varies with resolution and quality. The order of magnitude is a few cents to a few tens of cents per image, depending on the vendor and the tier you pick.
In our month: 12 single-image posts plus 4 carousels of 5 cards make 32 images. If each carousel needs two attempts on half its cards — and it will — add another 10. About 42 renders.
Here is where the first genuinely money-saving decision appears, and it is not about price: some of those images should not come from an image model at all.
A carousel card with a title, a number and a sentence; an offer card with a price and a CTA; a tip card with large type — all of that is typography on a background. Image models get text wrong, and you pay for the wrong attempt just the same. A rendered template solves those cases at a generation cost that is, literally, zero: it is HTML becoming a PNG in a headless browser.
Redoing the arithmetic with that split: 12 scene images come from the model, 20 carousel cards come from templates. The image line falls to less than half, and quality goes up — because the text comes out right.
Line 3 — video is the budget
Here is the line that decides everything. Generated video costs per second, and the difference between vendors and quality tiers is enormous — from a few cents per second to several dollars per second for the top models.
Four 8-second reels are 32 seconds of video. That sounds like nothing. But:
- Reels rarely work on the first attempt. Count on two to three takes per piece.
- That turns into 64 to 96 seconds generated for 32 seconds published.
With a cheap model, that is dinner. With a top-tier model, it is more than the subscription to any tool in this category. The same amount of content, with a two-orders-of-magnitude swing in cost — and the variable is a choice that, in most tools, you do not get to make.
What the arithmetic reveals
Add the three lines and the shape is always the same, almost regardless of current prices:
- Text: near zero. Ignore it.
- Images: small, and halves when you use templates for what is typography.
- Video: everything else. Practically the whole budget.
That changes what “controlling the cost” means. It is not picking the cheapest tool: it is being able to choose the video model, and being able to swap it when money gets tight or when a vendor launches something better at half the price. Anyone who does not choose the video model does not control the cost of their content — they only control the subscription.
Two traps that show up later
Regenerating should not delete. If every new attempt overwrites the previous one, you pay again just to compare. Keeping the previous takes — EverFeed keeps up to ten — turns comparison into a free operation. Trimming a video should not overwrite the original either, for the same reason.
Redoing everything when only one step is bad. If the caption came out great and the image did not, redoing the whole piece pays for the text again for no reason. That is why the EverFeed pipeline has one row per step, with the option to lock the good ones: a locked step survives any regeneration.
Subscription-with-generation versus subscription-plus-your-key
With the arithmetic above in hand, you can compare the two commercial models honestly.
A subscription that includes generation has a real advantage: predictability. You pay a fixed number and stop thinking about it. And it has two disadvantages: the price embeds a margin on a cost you cannot see, and the usage cap tends to arrive in exactly the month you most need to publish.
A subscription plus your own key inverts that: the monthly fee is smaller and predictable, the generation cost is variable and visible. In a slow month you genuinely spend less. In a busy month you spend more — but you see exactly on what, and you can swap the video model for a cheaper one without swapping tools.
For most of the small businesses we tested, the second shape works out cheaper because the video line — the one that dominates the arithmetic — becomes a choice. And because failed attempts acquire a price you know, which changes behaviour: people who see the cost per take learn to write better prompts remarkably fast.
How to do your own arithmetic in ten minutes
- Write down how many pieces of each type you publish per month. Be honest, not optimistic.
- Multiply the videos by 2.5 — that is the realistic attempt rate.
- Open the pricing page of the vendor you would use for video and multiply by the total seconds.
- Do the same for images, remembering to take out everything that is typography and could be a template.
- Add ten dollars for the text and move on — you will get change.
The number that comes out of that is your real content cost. Compare it with what you pay today. If you cannot do step 3 because you do not know which model your tool uses, that is also an answer.
Write your brand once.
With your own key, that arithmetic arrives itemised on the vendor invoice — and EverFeed charges nothing on top of it.
See both paths →