OpenAI’s ChatGPT Images 2.0 Actually Writes Text Now

OpenAI just dropped ChatGPT Images 2.0, and honestly, it’s about time they tackled the text rendering mess. The old model was notorious for generating images where any text looked like a toddler’s scribble—letters merging, words turning into gibberish. Version 2.0 fixes that, and the results are surprisingly good.

I’ve been testing it for a few days, and the improvement in text clarity is night and day. You can now generate images with proper English sentences, and even complex typography—like logos or signs—comes out clean. The model handles multiple fonts and sizes without breaking a sweat. This is huge for anyone who needs mockups, social media graphics, or just fun memes with readable text.

But the real surprise is multilingual support. I threw Hindi, Chinese, and Arabic at it, and it handled all of them reasonably well. The Arabic script, which has tricky ligatures, came out better than I expected. Not perfect—some characters got a bit squished—but leagues ahead of the previous version, which would just give up and spew random symbols.

The visual reasoning upgrade is subtle but noticeable. The model now understands spatial relationships better. Ask it to generate “a cat sitting on a table next to a cup of coffee” and it actually places the cat on the table, not floating in mid-air. It also handles multiple objects without mixing up their positions. This is higher than I expected from a single model update.

There are still quirks though. Complex scenes with lots of overlapping elements can confuse it. And if you push it with very long prompts, it sometimes drops details. But for everyday image generation, this is a solid upgrade.

One thing that bugs me: the pricing. This model is only available to ChatGPT Plus and Pro subscribers. Free tier users get a watered-down version with worse text rendering. OpenAI’s usual tiered access strategy, which I get from a business perspective, but it feels stingy given how many people just want to make decent-looking images without paying.

Also, the model still struggles with hands and fingers. I know, I know—it’s a meme at this point. But if you’re generating people, expect some extra digits or oddly bent fingers. Text is fixed, but anatomy isn’t.

Overall, ChatGPT Images 2.0 is a meaningful step. The text rendering improvement alone makes it worth trying if you’re a subscriber. It’s not going to replace dedicated tools like Midjourney or DALL-E 3 for professional work, but for quick, readable image generation, it’s now a viable option. Just don’t ask it to draw hands.