Ok fam, I spent Sunday afternoon doing something I've wanted to do for months.
I took the same 3 ad prompts and ran them through 8 different AI image models. Same brand (NordVPN), same words, same everything. One shot each, no do-overs, no cherry picking. Then I lined them all up to see who could actually make an ad I'd put money behind.
Total cost to run the whole thing: $2.02.
Two bucks. For 24 ad images. I've paid more than that for a bad coffee.
And I've gotta open with a confession...
Back in April (Build Notes #004) I gave you a JSON template for making product ads and told you the no_text_in_image: true line was CRITICAL. My exact words were "AI-generated text in images still looks terrible in 2026. Always add text in post-production."
Yea... about that.
Five of the eight models just rendered SIXTY words of exact ad copy with zero errors. Spelling, punctuation, line breaks, the brand name, a bold closer at the end. Perfect.
I was wrong. Four months ago I was right, and now I'm wrong, and that's how fast this stuff moves. Receipts below.
The 3 ads I made them all do
Quick setup so the pictures make sense.
Ad 1, the text ad. You know those hand-written Apple Notes ads that are all over feeds right now? The Grüns gummies one that starts "Sorry to whoever sees this and it's too late..." I stole that format and rewrote it for NordVPN. About 60 words of exact copy, no photo, no mascot. Just text. Hardest possible test for these things.
Ad 2, the product ad. Phone sticking out of a jeans pocket, breach alert on the screen, headline up top, brand name at the bottom. Straight-up copy of an ad NordVPN is running right now. Four separate chunks of text plus real-looking denim plus a believable app screen. All in one image.
Ad 3, the UGC ad. UGC is user-generated content, the ads that look like a customer shot them on her phone. Mine was a kitchen-table photo of a daughter setting up NordVPN on her mom's phone after mom's card got skimmed on vacation. Two faces, four hands, bank statement on the table. That's the "does it look like a real human" test.
All three prompts were written in JSON, same as I taught you in #004 and #005. (If you missed those, JSON is just writing your instructions as a labeled list instead of a paragraph, so the machine can't misread you.) That's the whole reason this test works. Eight models got the exact same instructions, not eight interpretations of a sentence.
Ok. Cheapest to most expensive. Here we go.
Photon 1 Flash (Luma) · $0.004 per image
Text ad: pure gibberish. Not one readable word. It nailed the little Notes toolbar at the top and then wrote alien handwriting underneath it lol.
Product ad: no headline, no brand name, and the phone screen says "The hceve Wedl tor waitmelr breach will k beam." 🤦♂️
UGC ad: honestly not bad. Real hands, two real people. But it went dark and moody when I asked for an amateur phone snap, and it stuck a NordVPN wordmark on the phone after I told it not to.
Four tenths of a penny gets you a decent photo and zero words. If you add all your text in Canva anyway, it's a real option for cheap backgrounds. For an ad with copy in it, no.
FLUX.2 [max] · $0.04 per image
Text ad: "WEY DO IN KIN COMGRER POST." That's the headline. Then eleven lines of nonsense. And then at the very bottom, in perfect bold letters... "NordVPN."
It knows the brand. It cannot write the sentence. I laughed out loud.
Product ad: made up its own headline, garbled the screen, and added a ® mark nobody asked for.
UGC ad: good! And it gave the mom a "WORLD'S BEST MOM" mug on its own, which is exactly the kind of detail that makes a photo feel real.
FLUX is a photographer, not a typesetter. Same as Photon at 10x the price.
Ideogram V4 · $0.06 per image
Text ad: PERFECT. All 60 words, apostrophes and all. Only nitpick is the frame looks more like a browser window than the Notes app. Nobody in a feed will care.
Product ad: all four pieces of text correct. The alert card came out small and the second line's a little faint, but it's right.
UGC ad: clean, real hands, decent scene. The daughter looks a little stern where I asked for warm.
Ideogram's known as the text specialist and it showed up like one. Six cents, and one of only two models in the whole test that let me pick a true 4:5 aspect ratio (that's the tall rectangle Meta's feed wants). Remember that.
Nano Banana 2 (Gemini 3.1 Flash Image) · $0.067 per image
Real quick because half the internet gets this wrong: Google's own docs say "Nano Banana" is FOUR different models. Nano Banana 2 is Gemini 3.1 Flash Image. Nano Banana PRO is Gemini 3 Pro Image, and it's coming up in a minute. Not the same thing, not the same price.
Text ad: flawless. Every word, the exact Notes toolbar, the bold closer. Of all eight text ads this is the one I'd run tonight without opening Photoshop.
Product ad: all four text targets clean, headline where it belongs, denim looks like denim.
UGC ad: two distinct faces, bank statement on the table, glasses and coffee where I put them. Reads a little polished, more stock photo than phone snap, but I'll take it.
Last Thursday I promised you a verdict on this one by name. Here it is: three for three, nothing to fix, under seven cents. Google calls it the workhorse and they're right.
Seedream 5.0 Pro (ByteDance) · $0.045 to $0.09 per image
Text ad: perfect copy, and the nicest paper texture and margins of the eight. Looks like a note somebody actually wrote.
Product ad: all four clean, and the best denim in the test. You can see the weave.
UGC ad: the BEST people of the whole test. Warmest light, most natural faces, closest anybody got to a photo a real daughter would've taken across a real table.
Three for three, beat everybody on the human shot, true 4:5, and it tops out at nine cents. Hold that thought.
Nano Banana Pro (Gemini 3 Pro Image) · $0.134 per image
Quick note on the price. Every other model in the picker shows a per-image price. This one bills per token instead (the little chunks of data these things count in), which is useless when you're trying to price a picture. So I just read the actual charge off my run log: 13.4 cents an image. That's what it really cost me.
Text ad: near miss. It dropped a letter ("and jus" instead of "and just") and rendered one "(I" in a totally different font. 98% right, and 98% right on an ad is 100% unusable.
Product ad: text is perfect. But it added the Nord shield logo next to the name after I told it in plain JSON no_logo_mark: true. The model decided it knew better than the brief.
UGC ad: excellent. Warm, natural, and it correctly labeled the paper on the table "Bank Statement," which nobody else bothered to do.
So the premium Google model cost exactly double the cheap Google model... and lost to it. Twice. If you've been paying for Pro because Pro sounds better, you've been buying the name.
GPT Image 2 (OpenAI) · $0.05 to $0.20 per image
Text ad: perfect. Line spacing runs a little loose but every word's right.
Product ad: the BEST product ad in the test. Biggest, most readable alert card, headline where it belongs. If a client sent me this I'd assume a designer made it.
UGC ad: very natural, great light. The bank statement's fine print says "Summary of Summary" which made me chuckle. Minor.
Three for three and the most consistent thing on the board. Also the priciest per-image model here, and it only offered 3:4, so you're cropping on upload. Earns its spot. I'm not sure it earns 20 cents.
Recraft V4.1 Pro · $0.25 to $0.30 per image
The most expensive model in the test. Up to 75x the price of Photon.
Text ad: perfect. Clean handwriting, no complaints.
Product ad: text is correct... and then it broke the brief twice in one image. I reserved the top of the frame for the headline, it put the headline at the BOTTOM. I said bold one word, it colored two words blue. Small stuff? Sure. But I paid the most for the model that listened the least.
UGC ad: weakest face of the eight. Mom drifts into that uncanny zone where the skin's too smooth and the eyes are too still.
The priciest model got the text right and everything else wrong. That's kinda the whole test in one collage.
Ok so who won?
Three went three for three: Nano Banana 2, Seedream 5.0 Pro, and GPT Image 2.
And look where they sit on the price ladder. The two cheapest models failed on text. The two most expensive models (Recraft Pro and Nano Banana Pro) both broke the brief. The winners are all in the MIDDLE, five to nine cents.
The floor failed and the ceiling didn't win. The money's in the middle. That's a lesson I've relearned about fifty times in twenty years and here it is again.
Also, credit where it's due: all EIGHT models made anatomically correct people. Not one mangled hand in 24 images. A year ago that UGC prompt would've broken half the field.
(One thing nobody nailed though. I asked for an amateur phone photo and every single model gave me polished ad photography. These things default to gloss. If you want ugly-real UGC you're gonna have to prompt for ugly on purpose. Future issue.)
My pick: Seedream 5.0 Pro.
It tied the other two on the scorecard, but it made the best humans of the eight, it gave me a true 4:5 with no cropping, and it tops out at nine cents when GPT tops out at twenty. Best people, right shape, under a dime.
Nano Banana 2 is a hair behind, and if you already live in Google's world it's a totally fine default. GPT Image 2 is worth the premium if you're doing complex product shots a client's gonna squint at. But if I've gotta pick ONE to bolt into the department, it's Seedream.
What this means for your wallet
Two weeks ago I showed you Motion's study of 578,750 ad creatives where only 4 to 8% ever become winners. So you need about 20 variants to find one.
Twenty variants from the winner costs $1.80.
For twenty years, creative volume was a budget question. You couldn't test 20 statics a week because designers cost money and have weekends. As of Sunday afternoon it's a HABIT question. The budget excuse died for two dollars.
Now go do this (it all connects)
This is a theme month, and the whole point is that the parts snap together. So here's exactly where today's winner plugs in:
1. Open the Ad Writer Agent from #027 and find the Generate Image block inside Images.flow. Set the model to Seedream 5.0 Pro. That's it. That's the whole change. Your ad writer now has an art department that costs $1.80 per batch of 20.
2. Go back to #004 and grab the JSON template idea. The three prompts I used today aren't test artifacts, they're your first three production templates. Text card, product-plus-headline, UGC candid. Swap the brand and copy, keep the structure.
3. Let the Daily Ad Brief from #025 babysit the batch. It reads your accounts every morning and applies the Wet Paint Rule (no verdicts inside 72 hours). Twenty variants a week is exactly what it was built to watch.
4. Flip the AI label switch in your ad accounts. Real quick on this one because it matters and it's new. Fifteen days ago (Aug 2) the EU's AI Act transparency rules kicked in. If you're running AI-generated ads to people in Europe, you have to label them so a human can tell. Fines up to €15 million. Google Ads already built the toggle (it applies in the EU, India, and New York) and Meta auto-labels under "About this ad." Two minutes to flip. I'm not a lawyer, but I know a free switch when I see one. Flip it.
So: copy from #027, images from Seedream, judgment from #025, label switch on... and Thursday we add the last piece.
I paid $2.02 to learn the tools crossed the line while most of us weren't looking, the expensive model isn't the good model, and the thing I told you in April is dead. Worth every penny.
Twenty years I've been telling people the pretty ad doesn't win. Turns out the expensive model doesn't either. 🤣
Sources cited in this issue
- MindStudio, model list and pricing (in-editor picker, verified 2026-08-17): https://www.mindstudio.ai/models
- Google, Gemini API "Nano Banana" naming (four distinct models): https://ai.google.dev/gemini-api/docs/nanobanana
- Motion, Creative Benchmarks 2026 methodology (578,750 creatives): https://motionapp.com/library/research/creative-benchmarks-2026/methodology
- European Commission, Article 50 AI Act transparency obligations (effective 2026-08-02): https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
- Google Ads Help, AI-generated content label setting: https://support.google.com/adspolicy/answer/17257106
- Build Notes #004 (April 2026), the
no_text_in_image: trueadvice this issue retires
Continue with Build Notes+
This Monday deep dive is free in full. For complete Thursday builds and supporting files, join Build Notes+.
