The 12x Spread: What AI Image Generation Actually Costs at Catalog Scale

The leaderboard is not the bill

Ask a room of designers which image model is best and you will get a confident answer, usually the one at the top of the arena. Ask the same room which image model their company should actually run at volume and the confidence evaporates. Those are two different questions, and in October the gap between them widened into something that looks almost like a market failure.

A pricing breakdown published this month put real numbers on the problem. For a team generating 50,000 product images a month at roughly 1K resolution, one output per request, the monthly bill across models that all consider themselves reasonable choices runs from about $840 to about $10,550. That is a 12.5x spread before anyone accounts for retries. Same workload, same resolution, same job.

The rate card, translated

The spread is easiest to see per image. At the cheap end, a fast batch model lands around $0.007 per 1K image. Nano Banana 2.1 sits at about $0.0168 in batch mode and $0.0336 in standard mode. FLUX.2 [pro] starts around $0.03, MAI-Image-2.6 around $0.039 by arena estimate, Grok Imagine Image 2.0 around $0.04, Seedream 5.0 Pro around $0.045, and FLUX 3 Image at 1K around $0.048.

Then the prices climb sharply. GPT-Image-2 at medium effort lands near $0.053, Nano Banana Pro near $0.134, and GPT-Image-2 at high effort near $0.211 per image. On the same 50,000-image workload, the difference between the cheapest sensible option and the most expensive is a house payment, or several.

The wildest part is how quickly resolution reshuffles the ranking. At 1K, the top few models are within spitting distance of each other. At 4K, the same 50,000 images cost about $3,780 on Nano Banana 2.1 and about $30,350 on FLUX 3 Image. A model that looked like a bargain at catalog resolution can become untenable the moment the brief asks for print quality.

Why cheaper often wins, and sometimes does not

The instinct when a model is "only two arena tiers behind" is to treat the difference as noise. For a product catalog, that instinct is usually correct. If your acceptance bar is "clean, on-brand, correct product", a cheaper model clears it more often than the leaderboard suggests, and the money saved goes back into shooting more angles or testing more concepts.

For a hero image that goes on a billboard, the calculation flips. The cost of a rejected image is not the generation fee. It is the designer's time, the reshoot, the missed deadline. When stakes are that high, spending four times more per image is cheap insurance, and the leaderboard stops being a vanity metric.

The pragmatic answer most teams are landing on is to stop picking one model. Run the cheap one on a batch of 50 of your own products, run the expensive one on the same 50, then count how many images pass first time. That number, not the arena score, is the one that decides your rate card.

There is a useful asymmetry hiding in that advice. Cheap models fail in obvious ways, and obvious failures are cheap to catch. Expensive models fail in subtle ways: a slightly wrong logo, a shadow that falls the wrong direction, a hand with one finger too many. Those failures cost more to catch because they survive the first glance, which is exactly why teams pay for the expensive model in the first place. The question is never whether the premium is worth it in the abstract. It is whether the specific mistakes the cheaper model makes are the ones your reviewers would have missed anyway.

Retries are the hidden multiplier

The single most under-accounted cost in AI image generation is the retry. The pricing model above assumes one output per request. Real workflows do not behave that way. If a team needs an average of 1.5 attempts per accepted image, the bill rises by 50 percent with no change to the unit price. Suddenly a cheaper model that needs two attempts costs more than a dearer one that needs one.

This is why "acceptance rate" deserves to sit next to "price per image" on any comparison sheet. A model with a 60 percent first-pass acceptance rate is effectively 1.67x its sticker price. A model with a 90 percent rate is effectively 1.11x. The gap between two models' effective prices can be wider than the gap between their list prices, and it points in a different direction.

Run the math on the two ends of the table and the lesson sharpens. A model at $0.02 per image with a 50 percent acceptance rate costs $0.04 per accepted image. A model at $0.05 per image with an 85 percent acceptance rate costs about $0.059 per accepted image. The "expensive" model is only half again as costly in practice, and it saved a human several rounds of review. Once you price the human time that goes into those retries, the gap can close entirely. The sticker price is a starting point, not an answer.

Deprecations are a line item too

October added a second kind of cost that does not appear on any rate card: the forced migration. Google scheduled the retirement of its earlier Gemini image endpoint for late October, which pushes anyone still on the previous generation to move whether or not they wanted to. On the workflow above, the move from the older endpoint to Nano Banana 2.1 cuts the monthly bill nearly in half and lands on a model that ranks higher. A free win, delivered by a deadline.

That is the good version of a deprecation. The bad version is the migration you did not plan for, on a model you cannot pin, at a price you did not model. If your image pipeline depends on a specific endpoint, the calendar is now part of your cost structure alongside the rate card.

The economics, not the aesthetics

None of this is an argument that cheaper models look just as good. They frequently do not. It is an argument that "which model is best" is the wrong question for anyone producing images at scale. The right questions are narrower and less romantic: what is my acceptance bar, what is my retry rate, what resolution does the deliverable actually need, and what happens to my bill when that resolution changes.

The 12x spread is not a bug in the market. It is the market telling you that image generation has matured enough to have tiers, the way cloud storage and video encoding did before it. The teams that treat image generation as a procurement problem, with acceptance rates and unit economics, will spend a fraction of what the teams that treat it as a taste problem spend. Both can be right about which picture is prettier. Only one of them can explain the invoice.