OpenAI’s Next Image Model Just Leaked. The Examples Are Insane.

GPT-Image-2 was hiding on LMArena under three code names. People found it. It already makes Nano Banana Pro look outdated.

Someone at OpenAI forgot to close the curtain

GPT-Image-2 was quietly live on LMArena under three aliases:

• maskingtape-alpha

• gaffertape-alpha

• packingtape-alpha

All three got pulled within hours of people identifying them.

That pattern means one thing. A release is coming.

---

What it can do

The range of outputs people shared in the window it was live was genuinely surprising.

Photorealistic portraits that multiple people could not identify as AI. Beach selfies with three people, natural lighting, correct hands, accurate sunglass reflections. The kind of output that would have been impossible six months ago.

Text rendering that actually works. The yellow filter problem from GPT-Image-1 appears to be fixed. Text sits inside scenes correctly instead of floating awkwardly on top of them.

World knowledge that shows. The model does not just know what things look like aesthetically. It knows what they look like specifically.

People generated:

• IKEA storefronts at night that passed as real photographs

• YouTube and Windows interfaces recreated accurately enough to be mistaken for screenshots

• Medical handwritten notes with realistic penmanship

• Minecraft scenes with correct in-game UI and art style

• Comic book panels with Spider-Man and Batman, readable speech bubbles, accurate costume details

• A fake “Claude Opus 5 Internal Document” inside a Minecraft world that went viral with 439K views

When asked directly what model it is, it claims to be OpenAI.

---

What people who tested it said

“This model is exceptional at realism and text. At long last DeepMind has proper competition. It also has better world knowledge than Nano Banana Pro.”

“Nano Banana Pro got cooked. OpenAI’s new image model just leaked and it outperforms NBP in everything.”

“Holy smokes. This is 100% AI.”

The consensus across every thread: this is a meaningful step above anything currently available publicly.

---

Why the removal matters

Three separate code names running simultaneously suggests OpenAI was testing multiple variants, probably with different safety or quality tuning, to see which performed best in blind evaluations before picking one to ship.

They pulled all three fast. That is not what you do with a model that is not ready. That is what you do with a model that is about to launch.

---

The competitive picture

GPT-Image-1 was a real step forward when it launched. It also had real problems. Yellow tint. Inconsistent text. Hands that still broke in complex scenes.

The early GPT-Image-2 examples suggest most of those problems are solved. And the people who compared it directly to Nano Banana Pro said it wins across most categories.

Nano Banana Pro was the photorealism benchmark for most of early 2026. If that has already changed before the model is even publicly released, the next few weeks are going to be interesting.

---

When GPT-Image-2 drops, most people will use it the same way they used GPT-Image-1. Here is how to not be most people.

The premium section below is the practical playbook. What the new capabilities actually unlock, the exact prompts that get the best output from photorealistic image models, and the specific use cases worth building into your workflow now.

Here is what is inside:

▫️ The prompt architecture for photorealistic output — the specific structure that separates outputs that look real from outputs that look AI

▫️ The text-in-image playbook — now that text rendering actually works, here are the use cases worth building around it

▫️ 10 copy-paste prompts — for portraits, product shots, UI mockups, editorial visuals, and more

▫️ The world knowledge use cases — the specific tasks where GPT-Image-2’s contextual accuracy changes what is possible for content creators and marketers

▫️ The workflow for newsletter and LinkedIn visuals — how to produce scroll-stopping images at scale without a designer

▫️ What to combine it with — how GPT-Image-2 fits into a broader AI content stack with Claude, Midjourney, and Canva

[Start your free 7-day trial →]

Cancel anytime. First subscribers get 50% off forever.

---

here we go🔥

1. The prompt architecture for photorealistic output