The best AI model for product shots, by cost per usable image

Three product briefs, five image models, three runs each. We counted which images we'd actually hand to a client and worked out what each usable one cost in credits.

Jaxson McNeilOctober 10, 2026

A square glass perfume bottle of amber liquid with a black cap and ASHE etched on the front, standing on wet black slate

The price on a model tells you what one image costs. It doesn't tell you what a good one costs, and for product work that's the only number that matters. If a model gets the bottle wrong two times in three, its real price is three times the sticker.

So we wrote three product briefs of the kind a small brand would actually send, ran each one three times on five models, and counted the images we'd be happy to deliver. That's 45 images. All of them are below.

How we picked the five

In our nine-model shootout, two models invented label copy nobody asked for and one turned coffee grey. We dropped those three. We also left out FLUX.2 [flex], the most expensive model in the studio, and Grok Imagine Image 2.0, the slowest. We kept FLUX.2 [max] in to see whether its price buys anything on product work. That left FLUX.2 [pro], FLUX.2 [max], GPT Image 2, GPT Image 2.5 Flare and Recraft V4.1.

What counted as usable

An image was usable if it matched every product detail in the brief and had nothing visibly wrong with it. Props got some slack: a third pebble where we asked for two didn't fail an image. A missing faceted base on a perfume bottle did, because the bottle is the product. We judged every image the same way and we've listed each call in a table under its grid, so you can disagree with us image by image.

Perfume bottle on wet slate

FLUX.2 [pro]'s image for this prompt, take 1 — Prompt
15 images
product photo of a square clear glass perfume bottle with a heavy faceted base and a matte black cap, the word "ASHE" etched small on the front, amber liquid inside, standing on wet black slate, a single softbox from behind making the glass edges glow, small water droplets on the slate, clean studio background
Try this prompt

FLUX.2 [pro]'s image for this prompt, take 2 — Prompt
15 images
product photo of a square clear glass perfume bottle with a heavy faceted base and a matte black cap, the word "ASHE" etched small on the front, amber liquid inside, standing on wet black slate, a single softbox from behind making the glass edges glow, small water droplets on the slate, clean studio background
Try this prompt

Model	Run 1	Run 2	Run 3
FLUX.2 [pro]	usable	usable	usable
FLUX.2 [max]	no	usable	usable
GPT Image 2	no	no	no
GPT Image 2.5 Flare	usable	usable	usable
Recraft V4.1	no	no	no

This was the hardest brief. "Heavy faceted base" is the kind of detail a designer cares about and a model can easily skip.

FLUX.2 [pro] and GPT Image 2.5 Flare got the facets in all three runs. Flare's cut-glass base and soft etched lettering looked like a real bottle. FLUX.2 [pro] took "softbox from behind" literally and put a glowing white disc behind the bottle every time. We liked it. Some art directors won't.

GPT Image 2 made the most beautiful perfume photographs in the whole test, with gorgeous light and a subtle, believable etch. It also gave all three bottles a plain thick base with no facets. That's 0 for 3, and it hurts. Recraft V4.1 skipped the facets as well and set "ASHE" large on the front or on the base, where we'd asked for small.

White sneaker, floating

Model	Run 1	Run 2	Run 3
FLUX.2 [pro]	usable	usable	no
FLUX.2 [max]	no	usable	usable
GPT Image 2	no	no	no
GPT Image 2.5 Flare	no	no	usable
Recraft V4.1	usable	no	usable

Every model made a believable shoe. The sole is where they split. We counted a sole as gum only if it read as gum from the side. A white midsole with a thin gum strip underneath didn't count.

GPT Image 2 went for a white midsole on all three runs and Flare on two. Both also gave us flatter, nearly side-on angles rather than three-quarter. Their outputs looked so alike that you could swap them and few people would notice.

The FLUX models and Recraft mostly got the full gum sole. FLUX.2 [max]'s first run had the bow's loops flopping over the side, not "neatly tied". FLUX.2 [pro]'s third run gave us the white-midsole look. Recraft's second run left the shoe sitting on the floor rather than floating above it.

We ran this brief once before with slightly different wording for the backdrop. The tally was the same for every model except GPT Image 2, which got one sole right that time. So these numbers aren't a fluke of one batch.

Skincare flat lay

Model	Run 1	Run 2	Run 3
FLUX.2 [pro]	usable	usable	no
FLUX.2 [max]	usable	usable	usable
GPT Image 2	usable	usable	usable
GPT Image 2.5 Flare	usable	usable	usable
Recraft V4.1	usable	usable	usable

The easiest brief, and four of the five models went 3 for 3. Every label in the set read "DAILY SERUM 30 ml" correctly.

The one miss was FLUX.2 [pro]'s third run, which put two dropper bottles in the frame when we asked for one. Flare again produced three near-identical layouts. That's useful if you need a consistent series and less so if you want options to choose from.

The numbers

Credits are what the studio charges per image. Seconds are the average time per image across all nine runs, with two requests in flight at once, so read them as relative, not absolute.

Model	Credits per image	Usable (of 9)	Credits per usable image	Avg seconds
GPT Image 2.5 Flare	7 credits	7	9.0 credits	12.7
FLUX.2 [pro]	30 credits	7	38.6 credits	8.0
GPT Image 2	21 credits	3	63.0 credits	17.4
Recraft V4.1	40 credits	5	72.0 credits	4.7
FLUX.2 [max]	65 credits	7	83.6 credits	18.5

Credits per usable image is simply the credits for all nine runs divided by the number we'd deliver.

Ranked verdict

1. GPT Image 2.5 Flare. Seven usable images out of nine. That's 9 credits per usable image, against 38.6 credits for the next best. It nailed the hardest brief. Its weakness is sameness. Its three runs of each brief came out close to identical, so rerunning it rarely gives you a different idea.

2. FLUX.2 [pro]. The same hit rate as Flare, with real variety between runs, and at 8 seconds the quickest model here after Recraft. If you want options to choose from, it's worth four times the credits.

3. GPT Image 2. When it followed the brief it made the nicest images in the test. It just didn't follow the brief often enough: 3 of 9, with none on the perfume bottle or the shoe.

4. Recraft V4.1. By far the fastest at under five seconds, and perfect on the flat lay. It ignored the faceted base and the "small" on the perfume, and lost one of three on the shoe.

5. FLUX.2 [max]. It matched FLUX.2 [pro] at seven usable images and cost more than double. We couldn't find a brief where it earned the difference.

If you shoot products, start on Flare. Switch to FLUX.2 [pro] when you need a different take and Flare keeps handing you the same one.

Try the perfume prompt on GPT Image 2.5 Flare