ChatGPT Images 2.0: Finally, Text That Doesn’t Look Like Garbage

OpenAI just dropped ChatGPT Images 2.0, and I have to say—this is the first time I’ve been genuinely excited about AI image generation in a while. Not because it makes prettier pictures (though it does), but because it finally fixes the thing that’s been driving me nuts: text that looks like it was written by a toddler with a seizure.

If you’ve ever tried to generate an image with a sign, a menu, or a book cover using the old model, you know the pain. Letters would bleed into each other, words would be missing, and anything beyond a single character was a gamble. The new model handles text rendering like an actual human designer. It’s not perfect—I still wouldn’t trust it with a corporate logo—but for social media graphics, posters, or even simple infographics, it’s shockingly good.

Multilingual support that actually works

The big surprise here is multilingual support. I threw some Chinese, Arabic, and Hindi text at it, and it didn’t just copy-paste random characters. It rendered the script correctly, with proper spacing and ligatures. That’s not trivial. Most image generation models treat non-Latin scripts as decorative shapes. This one seems to understand them as functional text.

I’m curious how far this goes. Can it handle mixed scripts? Right-to-left with embedded English? I haven’t stress-tested it yet, but initial results are promising. For global brands or creators working in multiple languages, this is a game-changer.

Visual reasoning: more than just pretty pictures

The “advanced visual reasoning” bit is what caught my attention. In practice, this means the model can generate images that require understanding relationships between objects. Things like “a chef holding a pizza box while a cat sits on a stool behind them”—it actually places the cat behind the chef, not floating in midair or merged into the pizza.

Is it flawless? No. I’ve seen it still mess up spatial relationships when things get crowded. But it’s a noticeable step up from the old model, where you’d get three hands and a dog where the stool should be.

What’s still missing?

I wish they’d addressed consistency across generations. If you generate the same prompt twice, you still get wildly different results. That’s fine for exploration, but annoying if you’re trying to iterate on a specific concept. Also, the model still struggles with fine details like fingers and small text (though the big text is much better now).

Pricing and availability haven’t changed—it’s still part of ChatGPT Plus and Enterprise tiers. No word on API access yet, which is a bummer for developers who want to integrate this into their own tools.

Should you upgrade?

If you’re already on Plus, you’ll get this automatically. If you’ve been holding off because image generation felt like a toy, this might be the update that changes your mind. The text rendering alone makes it useful for real-world projects. Just don’t expect it to replace a graphic designer for anything mission-critical.

I’ll be playing with this more over the next few days, especially the multilingual features. If you find any weird edge cases or impressive wins, let me know—I’m genuinely curious what this thing can’t do yet.