3 ad tests to run today with GPT Images 2.6
Three ad tests to run today with GPT Images 2.6.
In this issue
$10k in 1986 is ~$325k today, and 90% of pros couldn't do better

Hollywood loves a Wolf of Wall Street. In real life, the sheep in the index fund keeps winning.
The pros aren't the heroes here. Over time, 90% of professional investors fail to beat the index after fees.
And the index has receipts. Since 1986, the FTSE All-World has averaged ~9.1% a year through:
- The 1987 crash
- The dot-com bubble
- The 2008 financial crisis
- A pandemic
That turned $10,000 from 1986 into roughly $325,000.
The hard part is staying in. When markets drop, an index fund feels abstract. It's a lot easier to hold on when you know what you actually own.
So every Tuesday, 90 Percent pulls one random company from the FTSE All-World's 4,200 and tells you its story in 5 minutes.
No stock tips. Just the world you already own.
One good ad with AI? Easy.
But… a whole batch? You ask for a new background, and the model throws in a free rebrand. By the third ad, your logo looks like it belongs to a competitor.
ChatGPT Images 2.6 got plenty of attention on launch day for turning photos into GIFs. Two under-the-radar updates could come in just as handy:
- Saved image references let you reuse the same product photo across chats
- Edits run faster
Both help most when you need to tweak one thing and the clock’s ticking.
Here are three tactics you can put to work today.
GIF vs. still image: which gets more clicks?
Let’s turn one product photo into a looping GIF for an email.
The red perfume bottle stays put while clouds drift slowly behind it. All the movement happens in the background, so you can read the name on the bottle in every frame.
Now let’s see if those clouds do more than look good.

Send two versions of the same email.
One gets a still image, and the other gets a GIF of the same shot. Keep the subject line, copy, and CTA identical. Which version gets more people to click through to the product page?
Do it now:
- Turn the photo into a GIF: Upload the photo of the red perfume bottle against the clouds to ChatGPT and ask for a short, seamless loop. Have the clouds drift slowly from left to right while the bottle stays still and the framing stays fixed.
- Get the first frame right: Make sure the bottle stands out against the clouds and the label reads at a glance. Some email clients only show that first frame. Save it as the still image for your test, too.
- Run the A/B test: Put the first frame in version A and the full GIF in version B. Use the same subject line, offer, layout, and CTA in both. Randomly split your audience into two groups and send both versions at the same time.
- Compare product clicks: For each version, check what percentage of the people who received the email clicked through to the product page. Count each person only once.
Pro tip: Making a game? Create a quick GIF that shows how a character should move, then use it as a reference for a 3D model generator or Claude. That can cost less than generating a full video.
Show the same product in different settings
One photo of a perfume bottle can give you three different ads.

Take the red “tactics” bottle.
A spritz on the wrist makes it part of a daily routine. In an open purse next to a phone and keys, it’s a fragrance you can take with you. Put it against a dark backdrop with soft, blurred lights, and it feels like a night out.
Each shot helps people picture where the fragrance fits into their lives.
With Images 2.6, you can use the same saved reference for all three scenes, even if you create them in separate chats.
Then count how many come back without a surprise rebrand.
Playbook:
- Save the photo as a reference: Upload a photo of the bottle with the “tactics” label clearly visible. Ask ChatGPT to save it as a reference and name it Perfume.
- Create three shots: Type @Perfume in each new chat to pull in the reference. For the first shot, show a woman spraying perfume on her wrist. For the second, place the bottle in an open beige purse next to a phone and keys. For the third, set it on a dark, glossy surface with warm, out-of-focus lights in the background.
- See how many shots pass on the first try: Check the bottle’s proportions, ruby-red glass, and gold “tactics” lettering against the original. Check the spray nozzle in the wrist shot and the shape of the cap in the other two.
Pro tip: If the “tactics” lettering looks off, generate just the background and drop in a cutout of the bottle from your original product photo. This works best for the evening shot, where the bottle stands on its own.
See which image model fits your workflow

Let’s compare two image models in the OpenAI API: GPT-Image-2.6 Flare and Sunburst.
OpenAI designed Flare for fast iterations and Sunburst for precision in complex scenes. Which one gets you an approved ad faster, and for less? Give both the same brief and find out.
Take an Instagram Stories ad for HIGH TIDE, a fictional sparkling lemonade. The brief calls for a sky-blue can, a backlit wave, and a surfer riding a slice of lemon.
Say the wave looks perfect, but the can reads “HIGH TDE” instead of “HIGH TIDE.” You can’t run that ad, but it still cost you time and money.
The fastest drafts don’t always get you the fastest approval.
Step by step:
- Give both models the same text prompt: Send it to gpt-image-2.6-flare and gpt-image-2.6-sunburst through the API’s image generation endpoint, POST /v1/images/generations. Use the same 9:16 dimensions, quality setting, and level of complexity for both.
- Check the ad on your phone: Is the can the biggest and sharpest thing in the frame? Both lines of text on the can and the headline should be easy to read, with zero typos or extra text.
- Calculate the cost of an approved ad: Allow both models the same number of revision rounds. Track the time to the first usable draft and the time to final approval. Include every rejected version in your API costs.
The prompt we used
Create a photorealistic 9:16 Instagram Stories ad for HIGH TIDE, a sparkling lemonade brand. Place a matte sky-blue can covered in beads of condensation in the center of the frame. Behind it, a translucent wave curls into an arch. A miniature surfer rides the crest on a surfboard shaped like a slice of lemon [read more…]
Recap, logo intact
Each test ends with a number, so you’ll know what to keep and what to cut.
Move the clouds, measure the clicks: Put the GIF in one version of the email and its first frame in the other. See which version gets a higher share of recipients to click through to the product page.
One bottle, three ads: Use the same saved reference to show the fragrance in the morning, on the go, and after dark. Before you publish, check the bottle shape and make sure every letter on the label is still there.
The ads you scrap still cost money: Give Flare and Sunburst the same brief. Compare time to approval and cost per approved ad, including every rejected attempt.
If a model ever spells it HIGH TDE, at least you’ll know exactly what that typo cost.
The free Thursday newsletter that explains why customers actually buy

Sherlock had deductions. You get the science behind why people actually buy.
Every Thursday, PSYCHOLOGY OF MARKETING breaks down one proven principle and turns it into moves you can use.
Take a recent issue: simply asking customers about buying made them 3x more likely to open new accounts. Then: how to use it.
- Free, 5-minute read
- Real studies, real brands
- Tactics you can test this week
Readers agree. "Awesome insights with tangible info on how to apply," says Jenna.
CHATGPT: Can your checkout tell a shopping agent from a fraudster? OpenAI’s ChatGPT Wallet lets AI agents pay for users without seeing their card details, and every purchase needs a check like Face ID. It starts with US Pro subscribers. OpenAI’s playbook shows how to keep your firewall and fraud tools from blocking those sales.
AI AGENTS: …now need a permission slip at United. Its updated terms let search engines index public pages, but agents need written approval to search, book or buy. The rules also ban using its data to train or feed AI. Want that split on your site? United did it in its terms, not robots.txt.
AI AGENTS: Your Black Friday bot list just got shorter. In October, Meta’s and OpenAI’s agents made nearly 90% of agent requests that bot-security firm HUMAN tracked. Meta’s browsing agent, Muse, made 59% by itself. 85% of requests hit product and search pages, and only 1.9% reached checkout. Flag these agents to your bot protection team.
Get the next AI tactic in your inbox.
Ready-to-apply AI tactics delivered every Saturday to help you get wins on Monday. Consumed in 7 minutes or less.



























