We gave nine image models the same cold brew bottle

One prompt with a printed label, a hand and wet glass, run twice on every image model in the studio. Credits, seconds and what each one got wrong.

Jaxson McNeilOctober 10, 2026

A hand holds up a frosted bottle of cold brew with a white label reading North Pier Cold Brew, No. 07, in a blurred cafe

We wanted a prompt that would trip models up, so we packed four traps into one frame. A printed label with exact words on it. A hand holding something, with all five fingers in view. Wet, frosted glass. And a café crowd behind it all, out of focus.

Then we ran it twice on each of the nine image models you can pick in the studio. That's 18 images. Every one of them is below, unedited apart from resizing for the web, next to what it cost in credits and how long it took.

Going in, we expected the label to be where models fell apart. It wasn't.

The prompt and all 18 images

FLUX.2 [pro]'s image for this prompt, take 1 — Prompt
18 images
photo of a bartender's hand holding up a frosted glass bottle of cold brew coffee toward the camera, the paper label clearly printed with the words "NORTH PIER COLD BREW" and below it "No. 07", beads of condensation on the glass, all five fingers visible wrapped around the bottle, a busy cafe counter with several people out of focus behind, morning window light from the left, 50mm lens, shallow depth of field
Try this prompt

FLUX.2 [pro]'s image for this prompt, take 2 — Prompt
18 images
photo of a bartender's hand holding up a frosted glass bottle of cold brew coffee toward the camera, the paper label clearly printed with the words "NORTH PIER COLD BREW" and below it "No. 07", beads of condensation on the glass, all five fingers visible wrapped around the bottle, a busy cafe counter with several people out of focus behind, morning window light from the left, 50mm lens, shallow depth of field
Try this prompt

We asked for a 3:4 portrait frame. The three GPT Image models can't produce 3:4, so they returned their nearest tall shape (2:3). We sent Recraft V4.1 no shape at all and it came back at 896 × 1152 anyway. Keep that in mind when you compare crops.

Results

Times are the seconds between our script calling the model and the image coming back, averaged over the two runs. It's one evening on one server, so treat them as a rough guide. Credits are what the studio charges per image today.

Model	Credits per image	Avg seconds	Label exact
FLUX.2 [pro]	30 credits	8.9	2 of 2
FLUX.2 [max]	65 credits	19.1	2 of 2
FLUX.2 [flex]	75 credits	14.4	1 of 2
GPT Image 2	21 credits	16.4	2 of 2
GPT Image 2.5 Flare	7 credits	12.5	2 of 2
GPT Image 2.5 Sunburst	7 credits	17.8	2 of 2
Gemini 3.1 Flash Image	34 credits	11.5	2 of 2
Grok Imagine Image 2.0	40 credits	22.2	2 of 2
Recraft V4.1	40 credits	6.2	2 of 2

"Label exact" means both lines of the label match what we typed, character for character. It doesn't penalise extra text the model added on its own. That gets its own section below, because it turned out to matter more.

Text and hands are no longer the test

We built this prompt around the label because we expected exact text to be where models broke. In this run, 17 of 18 images printed "NORTH PIER COLD BREW" and "No. 07" exactly. The one miss was FLUX.2 [flex], which on its first run wrote "№ 07" with the numero sign. It's a typographer's substitution, arguably a nicer one, but it isn't what we asked for, and on a real product label that's a reprint. We didn't spot an extra or fused finger in any of the 18.

So if you're choosing between these nine for short label text or a hand holding something, our runs say you can mostly stop worrying. The differences that showed up were elsewhere.

The real difference: who adds things you didn't ask for

Look at the labels again. GPT Image 2.5 Sunburst printed our two lines correctly both times, then kept going. Run one added "SMOOTH. BOLD. BALANCED." and "12 FL OZ (355 ML)". Run two added "SMOOTH, BOLD, ALWAYS A GOOD DAY". Gemini 3.1 Flash Image did the same on its first run with "AUTHENTIC CRAFT" and a volume line.

For a mood picture that's harmless, even charming. For a label mockup it's a problem, because the client will read every word on that bottle and some of them will be words nobody wrote.

GPT Image 2.5 Flare and Sunburst also both put a smiling bartender's face in the frame and a chalkboard menu on the wall (with "matcha" spelled correctly, for what it's worth). We asked for a hand. The FLUX models, GPT Image 2, Grok and Recraft kept to what was in the prompt.

Model by model

FLUX.2 [pro], 30 credits. Both labels exact, both hands natural, condensation that looks wet rather than painted on, and the background properly soft as asked. It was also the second fastest at under nine seconds. This is the one we'd use.

FLUX.2 [max], 65 credits. Almost indistinguishable from [pro] on this prompt. We couldn't point to anything it did better, and it cost more than twice the credits and took twice as long.

FLUX.2 [flex], 75 credits. The best light of the lot: a low, warm window behind the bottle that makes the glass glow. It also gave us the only text miss. The most expensive model in the test didn't win it.

GPT Image 2, 21 credits. Exact text, sensible hands. One run read "bartender's hand" as a customer's point of view, arm coming in from the bottom of the frame. Plain and correct, with no surprises in either run.

GPT Image 2.5 Flare, 7 credits. The cheapest model here and a real bargain if you want a busy, lived-in scene. But it added a wave logo to our label on one run, and on both it staged a whole café around the bottle. Good for social posts. Not for anything where the label is the job.

GPT Image 2.5 Sunburst, 7 credits. Very similar staging to Flare, plus the invented label copy described above in both runs. We wouldn't use it for packaging.

Gemini 3.1 Flash Image, 34 credits. The second run was good, with a convincing pinch grip on the bottle neck. The first run turned the coffee a grainy grey that no one would drink and added label copy of its own. One good image from two.

Grok Imagine Image 2.0, 40 credits. Both runs look like a phone snapshot, which is either exactly what you want or not at all. The second run was the only image of the 18 where the glass itself was frosted, which is the literal reading of our prompt. It was the slowest, at around 22 seconds.

Recraft V4.1, 40 credits. The fastest at about six seconds and the most like an unretouched photo, with real-looking grain and messy natural light. The label text was correct but set small on a single line, so it reads less like a designed product.

Verdict

For a product held in a real scene, where the label has to say exactly what you typed, use FLUX.2 [pro]. It matched the prompt in both runs, made no additions, and cost 30 credits in under nine seconds. Nothing more expensive beat it here.

If you need volume and can live with a model that dresses the set for you, GPT Image 2.5 Flare at 7 credits gives you four images for less than the price of one from [pro].

Try this prompt on FLUX.2 [pro]