OpenAIs New Image Generator Finally Does Text Without Typos. Just Not in Your Language.

OpenAI's new image generator finally does text without the typos. Just not in your language.

ChatGPT Images 2.0 shipped Tuesday, and the headline upgrade is real: the model renders English text legibly. Menus, UI elements, signage, infographics — the things that made AI image generators useless for professional creative work are now, at least in English, actually readable. Ask it for a Mexican restaurant menu and it produces something you could hand to a server. Two years ago, the same prompt would have invented dishes like "enchuita" and "margartas." TechCrunch

The company declined to say whether Images 2.0 uses a diffusion architecture or an autoregressive approach. That distinction matters to researchers, but OpenAI wouldn't confirm it in a press briefing. TechCrunch The model does use the same "thinking" capabilities as GPT-4o — it can search the web, generate multiple images from one prompt, and double-check its outputs before returning them. TechCrunch Resolution goes up to 2K. You can generate images from 3:1 wide to 1:3 tall aspect ratios. WIRED

For English-language workflows, this changes things. Teachers making study guides, marketers generating social assets, product teams prototyping UI — all of them can now get legible text in the output without a human cleanup pass. That's the productivity gain OpenAI is selling, and independent testing by WIRED bears it out for English.

The catch is the other six thousand languages.

WIRED asked the model to generate a Chinese fan-poster collage and then asked it to translate its own output. The response was frank: "A lot of it is fake, or semi-gibberish AI text dressed up to look like Chinese meme-poster writing, so it does not all cleanly translate." WIRED The model produced characters that resembled Chinese script but didn't correspond to actual words. OpenAI has improved non-Latin text handling compared to earlier models, but the improvement is incomplete — the model itself admitted this when pressed.

This is not a minor gap. OpenAI has explicitly framed ChatGPT as a global personal assistant. CNET reports the company killed its Sora video product last month to concentrate on enterprise "core products" — and Images 2.0 is a central piece of that super-app strategy. CNET A global assistant that works in English but produces gibberish in Chinese, Hindi, Arabic, or Japanese is not yet global. OpenAI said it has improved non-Latin language support, but what "improved" means in practice is apparently still very uneven.

The model has a knowledge cutoff of December 2025, which matters for anything involving recent news or events. The gpt-image-2 API is live, with pricing tied to output quality and resolution. All ChatGPT and Codex users get access; paid users get the more capable generation modes. OpenAI

The competitive pressure is real. Midjourney has dominated artistic image generation; Adobe Firefly has been the professional design workflow choice. Both have struggled with text. If ChatGPT can now produce on-brand marketing assets with correct legible copy from a single prompt, some fraction of that workflow migrates. The question is whether the English-only text reliability is enough to pull professional users away from tools they've already integrated. For now, probably not for everyone. But the direction is clear.

OpenAI has closed the gap it needed to close for English. It hasn't closed it everywhere else.