ChatGPT Images renders legible in-image text

Image generation has always had an annoying blind spot. It could make gorgeous scenes, but the moment you asked for a menu, a labeled diagram, or a poster with real words, things fell apart. Letters warped. Spacing broke. Non-Latin scripts got especially messy. That is the gap OpenAI is trying to close with ChatGPT Images 2.0, released on April 21, 2026, with a new model that leans hard into usable text inside images. (openai.com) ### What actually changed? The big change is that OpenAI is no longer framing image generation as just an art toy. The new release explicitly pitches improved text rendering, multilingual support, and stronger visual reasoning as headline features. In plain English, the model is supposed to do better at images where the words are the point, not decoration. That includes charts, inst(openai.com) concepts. (openai.com) ### Why was text so hard before? Because image models do not “write” the way a text model writes. They generate pixels, and text is brutally unforgiving at the pixel level. A slightly wrong hand still looks like a hand. A slightly wrong letter looks broken instantly. Dense layouts make that worse — every character has to be shaped correctly, spaced correctly, and placed in the right(openai.com)h GPT-4o image generation, but ChatGPT Images 2.0 is the clearer product push around that capability. (openai.com) ### Why does multilingual text matter? Because “legible text” is not just an English problem. If a model can only fake Latin characters convincingly, it is still weak for global use. OpenAI is calling out multilingual support directly in the launch materials, which suggests the company sees this as part of the core upgrade, not a side benefit. That is important for anyone making ed(openai.com)ups for more than one language market. (openai.com) ### Is this just for ChatGPT users? No — OpenAI shipped the same broader push into the developer stack too. The API docs point developers to GPT image models including gpt-image-2, and OpenAI’s developer announcement describes that model as its most capable image generator yet, with better layouts, editing, instruction-following, and text rendering. So this is not just a nicer button in ChatGPT. It is meant to be something products can build on. (developers.openai.com) ### Why do layouts matter as much as letters? Because most useful images are structured. A restaurant menu is not just a list of words. A diagram is not just labels. The model has to understand hierarchy, alignment, grouping, and visual logic. OpenAI keeps pairing “text rendering” with “advanced visual reasoning,” and that pairing is the real tell. Basically, the co(developers.openai.com)for, not just drawing prettier fonts. (openai.com) ### Does this replace manual design work? Not completely. The catch is that “much better” is not the same as “perfect every time.” Brand-critical work, long documents, regulated labels, and anything that needs exact copy still needs human checking. But the threshold has moved. If the first draft already gets the sign, chart, or explainer graphic mostly right, the job shifts from reb(openai.com)it sounds. (developers.openai.com) ### Why is this a bigger deal than prettier images? Because it makes AI images more useful for information, not just illustration. The old failure mode was obvious: great background, nonsense words. The new push is toward images you can actually publish, present, prototype, or hand to a customer with fewer fixes. That moves image generation closer to a real producti(developers.openai.com)by hand. (openai.com) ### Bottom line The headline is not that AI can finally make nicer posters. It is that OpenAI is trying to make generated images legible enough to carry information. If that holds up in everyday use, the most important upgrade is not aesthetic. It is operational. (openai.com)