Image 2.0

For two years, one thing gave away every AI-generated image. The text.

Posters with garbled menu items. Thumbnails with a letter missing from the title. Ads where the brand name had an extra vowel. You could spot a DALL-E output at 20 paces because the words on it were nonsense.

OpenAI shipped Image 2.0 this week. Text rendering is fixed. That matters more than it sounds.

Why I care about this one

I build thumbnail generation systems as part of my day job. One of the projects I’ve been working on for the last few months is Thumbnail OS. The biggest blocker has always been the same thing. The title on the thumbnail has to be readable. AI couldn’t do it. So the pipeline ended with a human fixing the text, or swapping the image model for a compositor that rendered the text on top.

That workflow is now optional. You can prompt a full thumbnail, text included, and get something shippable.

What’s actually new

Three things matter.

Text rendering. OpenAI showed a restaurant menu side by side with DALL-E 3. The old one said “enchuita” and “churiros.” The new one reads like an actual menu. Non-Latin scripts also improved, so Japanese, Korean, Hindi, and Bengali render properly.

Thinking mode. Images 2.0 is OpenAI’s first image model that reasons before it draws. It can search the web, plan the layout, batch multiple images, and verify its own output. Instant mode runs on every tier, including free. Thinking mode is behind the paywall.

Batch consistency. You can generate up to 8 images from a single prompt with the same character, same object, same style across the set. That’s what unlocks comic strips, storyboards, multi-panel campaigns, and product shots at different angles.

Resolution goes to 2K. Aspect ratios span 3:1 to 1:3. DALL-E 2 and 3 are retired on May 12.

What this changes about how you work

If you build anything that outputs images at scale, you need to revisit your pipeline.

Any step in your workflow that exists because the text was wrong: you might not need it anymore. Any step where a designer had to redraw the title, add the caption, fix the menu line: reassess.

The second change is batch jobs. Eight consistent images from one prompt means campaign work that used to take a day in Figma or Photoshop can run in one shot. You lose the fine control. You also lose the fine drudgery.

If you’re on a creative team, this is the moment to sketch what your workflow looks like if image generation becomes reliable. Not in a year. Now.

Where this falls short

Thinking mode takes minutes per image. If you’re used to the instant generation loop, this is a different rhythm. Batch and walk away.

Knowledge cutoff is December 2025. Anything recent, like a logo redesign or a product launch from this year, won’t be accurate without reference images.

OpenAI won’t share the architecture. You’re betting on a black box, and that box is owned by one vendor. If your company runs on Gemini or Midjourney for production creative, this is another model to evaluate, not a drop-in.

And “surprisingly good at text” is not the same as “reliably good at text.” Expect failures on dense copy, brand-specific fonts, and tight kerning. Always proof the output.

Where this lands

Text on images was the single thing that kept AI creative out of production. That gate is gone, or at least half gone.

If you ship visual assets at scale, the next 12 months are going to look very different from the last 12.