Doubling both edges quadruples a per-megapixel bill
Print-resolution models meter area, so 1024² and 4096² are the same prompt at roughly 16× the credits.
Print-resolution models meter area, so 1024² and 4096² are the same prompt at roughly 16× the credits.
That is the only square law on the cost hub. Every other meter is linear: twice the seconds, twice the credits. A canvas twice as wide and twice as tall is four times the pixels, so four times the charge. Three doublings — 512 to 1024 to 2048 to 4096 — is a 64× bill. Buyers underestimate this by exactly that factor, because "bigger image" sounds like a quality toggle, not an area meter.
Area is the formula
Credits = rate_per_megapixel × (width × height ÷ 1,000,000). Both edges enter. Aspect ratio only matters through area: a 16:9 and a 9:16 of the same pixel count cost the same; a wider frame at the same height costs more.
Thirty-three image models bill this way. Rates run from 1 credit a megapixel (Ernie Image Turbo, Flux 2 Flash, Flux Schnell) to 15 (LongCat Edit). The worked example on the cost page uses Flux 1.1 Pro at 4 credits a megapixel:
- 1024 × 1024 ≈ 1.05 MP → about 4 credits
- 2048 × 2048 ≈ 4.19 MP → about 17 credits (4× the previous row)
- 4096 × 4096 ≈ 16.78 MP → about 67 credits (4× again, ~16× the 1024 square)
Same prompt. Same model. The card that shows "4 credits" is the rate, not the bill. Ask for more canvas and the charge scales. The app quotes the whole-credit total before the render runs. The scenario page is print-resolution AI image cost.
1024² versus 4096² is the same prompt
The expensive mistake is treating print scale as free detail. On a flat-billed image model you should ask for the largest output available — size is free. On a per-megapixel model, asking for a poster when you needed a thumbnail is the most expensive version of the same prompt on offer. A discarded print-scale render still costs the area price. Test the prompt on a social square, then commit to 4096.
Text-to-image is the generate surface. It does not warn you that you just quadrupled the bill by ticking both edges. You pick the model and the size. The formula does the rest.
A vertical story frame (1080 × 1920) and an HD landscape (1280 × 720) differ in shape but barely in area, so they cost nearly the same at a given rate. A 4K landscape (3840 × 2160) is several times either. Shape is a composition choice. Area is the invoice.
The other two image meters do not do this
Not every still is area-priced. Flat per-call image models ignore dimensions entirely — one price per image, so the largest canvas is the rational default. A handful of image rows use resolution_flat: one price up to HD, a higher price at 4K. That is a two-step band, not a continuous meter. Every size inside a band costs the same; crossing the band is a cliff.
Only the per-megapixel set scales continuously with area. If you are comparing a Flux 1.1 Pro print square to a flat-billed catalog shot, you are comparing two formulas, not two "image prices." What 100 AI product images cost is the flat-call book. Do not mix the pages.
Upscaling is a third charge, on a third model. Generating small and upscaling is often cheaper than rendering large in one pass — or it is not, depending on the upscaler. Upscale an image's resolution is its own tool (upscale_image). Work out both totals. Do not assume "upscale later" is a discount.
Print production itself is not in the credit. Credits buy the file. Ink, stock, and a RIP are downstream.
Test small, print large
A practical loop: lock the prompt at 1024² (or the smallest canvas that still shows the error you care about). When the composition holds, raise both edges once — 2048 — and look at type, seams, and packaging type. Only then go to 4096 if a print spec actually needs that pixel count. Each step is 4× the last on this meter. Skipping the middle step is how a prompt-debug session becomes a print-scale invoice.
Batch volume compounds with area. Ten large canvases is ten times a large canvas. There is no volume tier. If the job is a catalogue of SKUs, the cheap path is a cheap rate and a canvas no larger than the listing photo. Save 4096 for the one hero that will be printed.
The figure on the model card will keep looking like a per-image price. It is not. It is a rate per megapixel. Read it that way or the square law will surprise you twice: once at 2048, again at 4096.
FAQ
Why does the model card show one credit figure?
On a per-megapixel row that figure is the rate, not the bill. Flux 1.1 Pro lists 4 — that is 4 credits per megapixel. A 4096 square is about 67. The confirm step quotes the whole-credit total for the canvas you actually asked for.
Is a wide image more expensive than a tall one?
Only if it contains more pixels. Width × height is the whole formula. Shape is free; area is not.
Should I generate large, or generate small and upscale?
Work out both. A large render is one area-priced charge. Small-plus-upscale is a small area charge plus a separate upscale on another model. Which wins depends on that upscaler, not on a rule of thumb.
Do all image models charge more for large output?
No. Flat-billed models ignore size. Resolution-banded models have two prices, HD and 4K. Only the thirty-three per-megapixel models scale continuously with area.