Here is the setup every pricing page runs. It shows a plan, a credit balance, a rate per image. Twenty dollars a month, five hundred credits, four cents a render. The math looks like buying prints. You picture a stack of finished images and divide. The trouble is that an AI render is not a print. It is a draw from a slot machine that mostly returns pictures you will delete, and the pricing page is quoting you the cost of one pull, not the cost of a jackpot.
What you buy on the way to a usable image is a pile of rejects. Wrong material on the facade. A window the model invented. Scale that slipped so the door reads at eight feet. A mood that is fine but not the one the client asked for. Each of those cost you a generation, and each one goes in the bin. The credit rate is real, but it prices the pull. Your project pays for the keeper.
The number that actually bills you
Write it as one line and the whole comparison changes. Your true unit cost is the price per generation multiplied by the number of generations it takes to reach an image you would present. Call that second figure the keeper rate. One in three means you burn three pulls to land one you keep. One in eight means eight.
Once you see it that way, the sticker price stops deciding anything on its own. A tool at four cents a render with a keeper rate of one in eight costs you thirty-two cents per usable image. A tool at nine cents a render that lands one in two costs eighteen. The dearer credit is cheaper where it counts, by a wide margin, and no pricing page on either site will tell you so, because neither one publishes a keeper rate. They cannot. It depends on your inputs, your prompt discipline, and how strange your building is, none of which the vendor controls.
The credit rate prices a guess. The keeper rate prices the answer. You compare tools on the first and you pay them on the second.
This is why the cheapest plan so often feels expensive by Friday. You signed up on the four-cent number, then spent the week rerolling a canopy that would not sit right, and the credit balance drained three times faster than the arithmetic promised. Nothing was broken. You were paying the real price the page never quoted.
A worked example, same budget, three tools
Say you need twelve presentation images for a scheme review and you have the same forty dollars to spend on each of three tools. Watch what the keeper rate does to it.
| Tool | Price per render | Keeper rate | Cost per usable image | Usable images for $40 |
|---|---|---|---|---|
| The cheap credit | 4¢ | 1 in 8 | 32¢ | about 125 |
| The mid plan | 9¢ | 1 in 2 | 18¢ | about 222 |
| The premium seat | 15¢ | 2 in 3 | about 23¢ | about 175 |
The tool with the highest sticker price is not the most expensive per keeper, and the cheapest sticker is the worst value on the board. The point is not the exact figures, which will move for your work. The point is the ranking flips the moment you stop counting pulls and start counting keepers, and the sticker price alone will lead you to the wrong tool more often than not.
What drags a keeper rate down
The good news is that the keeper rate is mostly yours to move. Four things push it into the floor, and all four are inputs you control.
A vague prompt. If the model is guessing at time of day, material, and mood, it will hand you a different guess every pull, and most of them miss. Every specific you pin down is a reroll you do not pay for later.
No control pass. Feeding a flat view with no depth or edge guidance lets the model rebuild your geometry from scratch, so walls wander and the keeper rate collapses. A depth or line pass locks the structure and turns rerolls from a coin flip into a nudge.
Asking for a hero and a set at once. Landing one striking image is a different job from landing six that match. If you judge every pull against a whole consistent set, your keeper rate falls off a cliff, and that cost belongs to the set, not the tool.
Unusual geometry. Odd cantilevers, tight interiors, anything the training data thinly covers, all raise the reroll count. This one you cannot fully fix, but you can know it going in and budget the extra pulls instead of discovering them mid-deadline.
Measure your own in one project
You do not need the vendor to publish a keeper rate. You can measure yours in an afternoon on a job you are already doing. Pick one real view. Render until you have an image you would actually put on the sheet. Count every generation it took, including the near-misses. That count is your keeper rate for that tool on that kind of work, and it is worth more than any comparison table, because it is measured on your building and your standards, not a demo scene tuned to make the pull look easy.
Do it once per tool you are weighing during the trial, before the credits run out, and you will have the only figure that lets you compare plans honestly. Log it next to the price and divide. The plan that wins on cost per keeper is the plan to buy, and it is frequently not the one that wins on the sticker.
Our take
Stop shopping on the per-credit rate. It is the one number every tool is happy to show you precisely because it flatters the tools that miss the most, and it says nothing about whether the image ships. Track cost per keeper instead, measure it yourself on one hard view, and carry that figure into every plan comparison you make this year. The pricing page is selling you pulls. Your practice bills for keepers. Buy the tool that gets you there in the fewest tries, and let the cheap credit rot on the shelf where it belongs.
Written from the 11 August 2026 intel sweep, which surfaced several 2026 AI render roundups leading with monthly prices and per-credit rates. ArchiGen AI carries no sponsored placements and runs no affiliate links in its rankings.