Introducing ChatGPT Images 2.0

OpenAI has quietly rolled out a new image model inside ChatGPT, and the pitch this time isn't just "prettier pictures." It's about getting the boring stuff right. Text rendering — the thing that made AI art look uncanny for years, with melted letters and gibberish signage — has apparently been overhauled. If you've ever asked a model for a birthday card that says something specific and gotten alphabet soup instead, that's the exact failure OpenAI says this update targets.

Multilingual support is the second big claim. Earlier generations of image models were trained overwhelmingly on English captions, which meant non-English text in generated images often came out mangled or simply wrong. ChatGPT Images 2.0 is being positioned as far more capable across languages, which matters for a huge chunk of ChatGPT's global user base that has never been well served by English-first tooling.

Then there's the visual reasoning piece, which is the more interesting technical bet. Rather than just pattern-matching pixels to a prompt, the model is described as reasoning about a scene before rendering it — working out spatial relationships, object counts, and logical constraints instead of guessing and hoping. That's the kind of capability that separates a novelty toy from something people actually use for design work, mockups, or diagrams.

None of this is revolutionary in isolation. Competitors have chased better text rendering and multilingual output for a while now. But bundling all three improvements into the default ChatGPT experience, rather than a separate product, is the part that will actually change how many people encounter it. Most users won't read a model card. They'll just notice the images finally look right.