You needed a product shot. Cotton henley, window light, the knit has to read as cotton, not as a plastic suggestion of fabric. You pick the cheapest name on the list and generate. Vinyl. You generate again. Then again. Five attempts later you have a frame you can keep, and you tell yourself each roll was cheap so the session was cheap.
It was not.
The number on the picker is the cost of an attempt. The number that hits the job is the cost of a result you can ship. Those are different units. Confusing them is how a fifteen-credit model becomes the expensive one.
The sticker price is not the bill
Per-generation price is easy. It sits next to the model name. It is the wrong unit for any job with a keep-or-discard rule.
A generation you cannot use is not a discounted asset. It is a spent attempt. Collapsed hands, a logo that almost reads, cotton that reads as plastic: billable, not usable. Reroll and you pay again. Repair it in an editor and you pay in time. Most sessions are a pile of near-misses and one frame that survives review. Pricing the pile as if it were the keeper is how the cheapest line item starts to look like the cheapest workflow.
On PicX Studio the per-image credit costs include Seedream 5 Lite at 15 credits, Seedream 4.5 at 18, Nano Banana 2 Lite at 20 flat, Nano Banana 2 at 35 for a 1K render, and GPT Image 2 and Nano Banana Pro at 53. Every model bills from the same single credit balance. Switching is free. Regenerating is not.
Cost per accepted result is credits spent, divided by images you actually kept. Acceptance rate is passes over attempts on a given kind of job. Divide the per-image price by that rate. Cheap attempts with a low keep rate are not cheap.
one API comparison framed the question as cost per accepted image, then argued that you should retain the path that clears your own keep test instead of treating a published ranking as the decision. Your brief. Your constraints. Your keep rule.
Do not invent a rate. Measure one. Take a brief you actually have. Same prompt, same references, same size. Run the cheap model and the dearer one you would otherwise avoid. Count generations. Count keepers. A keeper is something you would send. If you painted out a fifth finger, that was not a pass.
A handful of attempts is not a scientific sample. It is enough to catch an obvious mismatch. The winner is the lower cost per keeper, not the lower sticker. Acceptance is not a property of a model. It is a property of a model on a brief. Re-measure when the constraint changes.
The cheap model is often the expensive path
Seedream 5 Lite at 15 credits looks like the obvious default. Five attempts is 75. Nano Banana Pro at 53, if you keep the first one, is 53. The arithmetic is not subtle.
This is a worked example, not a claim about how often each model lands. Sometimes Lite lands first. Then it really is cheaper, and you should use it. Sometimes it does not. Then you have been paying a retry tax and calling it thrift.
Each click feels small. Fifteen credits. Twenty. One frame is almost right. You chase the almost. By the time you keep something, you have spent more than the model you refused because it sat at 53.
There is a second way cheap becomes expensive. You accept a worse result because "good enough" starts to look like a keep. The honest test is whether you would have kept this frame if it had been the first output from the dearer model. If not, you did not save. You lowered the bar.
We keep hitting this on hard constraints. A face that has to be the same person. Type that has to be the type. Cheap models will give you a beautiful almost. Almost is not a keep. Ten banners that must look like one campaign fail the same way. One frame glossy, the next painterly, the third with a different white point. Rerolling cheap until the set matches can cost more than starting dearer. It depends, and it depends on whether the job is a set or a single frame.
When the dearer image is the cheaper job
Some jobs have a failure cost that dwarfs the credit gap.
Client work is the obvious one. A hero that has to survive a review meeting is not a thumbnail. If the knit has to read as cotton and the cheap model keeps handing you vinyl, you are not exploring. You are late. The 53-credit model that lands a usable photograph is cheaper than an afternoon of 15-credit misses.
A face that must hold is the same shape of problem. A person who is almost the talent is not usable, even if the lighting is good. A product orbit can ignore this. A person cannot. Spend the credits on the model that holds the face.
Text inside the image is the third. A lockup, a label, a price, a packing line. If the letters have to be the letters, a pretty image with the wrong word is a miss. Patching type later is sometimes right. Sometimes the type is the picture. Then paying more per generation is straightforwardly correct.
It depends, and it depends on the constraint you cannot relax. If you can put the type on in an editor, do not pay a premium to roll dice on spelling. If the type has to live in the scene, catching the same light as the rest of the frame, the dearer model is often the cheaper path. Seedream 4.5 at 18 versus Nano Banana 2 at 35 for a 1K render versus GPT Image 2 at 53 only looks large if you pretend every generation is a keep. On a hard pass or fail, retry count dominates.
When paying more is waste
People burned by cheap rerolls overcorrect.
Exploration is not a keep-or-discard job. You are looking for a direction: a palette, a camera height, a wardrobe, a room. A miss is information. Paying 53 credits to learn that a wide shot is the wrong shape is waste. Run the cheap model. Generate a handful. Throw most of them away on purpose.
Thumbnails are the same. They have to read small. They do not have to survive a crop-in on the cotton. A Lite model that gets composition and color is doing the job. You do not need Nano Banana 2's 1K render for a grid of options.
Mood boards even more so. A board is a conversation about a direction. Vibe coherence matters more than material accuracy. Use the 15-credit and 18-credit names. Switch up only when a tile has to become the actual asset.
The tell is whether a near-miss still did the work. On a mood board, an almost is often enough. On a client hero, it is not. Name the phase before you pick the model.
Search cheap, lock expensive
Search on the cheap model. Seedream 5 Lite at 15, or Seedream 4.5 at 18, or Nano Banana 2 Lite at 20 flat. Same brief, loose enough to try camera, crop, wardrobe, light. Keep the ones with the right bones: pose, room, product angle. You are not making the final. You are finding it.
Then lock. Take the search frame you would actually send and rerun it on the dearer model: Nano Banana 2 at 35 for a 1K render, or GPT Image 2 or Nano Banana Pro at 53. Change the model, not the brief. Switching is free in credit-balance terms, so the extra cost is the dear generation plus whatever you still reroll at that tier.
This is a trade-off. You will sometimes search into a composition the dearer model does not reproduce. That is irritating. Then you choose: keep the cheap frame if the surface will take it, or spend a couple of dear attempts to recapture the bones. Do not run the entire search at 53. Do not skip the lock on a job that needed it either. Faces, type, fabric. Campaign hero: pay for the lock. Thumbnail: stop at search.
Cut width before you cut quality
When you need to save, people reach for the cheaper model, a lower quality setting, or fewer attempts. The lever that actually moves file size is width.
A generated PNG is large. A measured 1408x768 output came to 1697 KB. Bounding the delivered width does more for size than lowering quality on a too-wide image. Quality reduction smears the thing you paid to get right: the knit, the type, the edge of a face. Width reduction throws away pixels you were not going to show.
Match the output to the surface. A thumbnail does not need full width. Ask for the width you will deliver. Then, if you still need to save, pick the cheaper model for the phase you are in. Wide feels like quality. It is often just a large PNG.
Pay for the constraint that can fail the job. Stop paying for the ones that cannot. Per-image price is a line item. Cost per accepted result is the bill. Measure the second. Pick the model for the phase. Cut width before you cut the thing the image had to get right.
