OpenAI has launched ChatGPT Images 2.0 for ChatGPT and Codex users worldwide, with a more powerful compute tier available to paid subscribers. The new image model focuses on two areas that have long limited AI image tools: reasoning and text rendering. It can also generate multiple images from a single prompt and supports aspect ratios ranging from 3:1 to 1:3.
A single prompt can produce a full sequence of images
Compared with earlier models, Images 2.0 brings in stronger reasoning capabilities. According to the released details, the model performs more thinking steps before generating an image and can search the web for current information. Its base knowledge cutoff is listed as December 2025. That gives it more room to handle structured visual tasks rather than isolated one-off outputs.
One example cited in testing involved an infographic for San Francisco’s weather forecast and suggested activities for the next day. The model was able to combine weather details with local landmarks such as the Ferry Building, Castro Theatre, and the Transamerica Pyramid in a single composition.
English text looks much cleaner than before
Text inside AI-generated images has been a weak point for years. Letters warp, words drift, and spelling often collapses under more complex layouts. In outside tests, Images 2.0 showed a clear jump in English text rendering, producing words that appeared far sharper and more accurate than earlier generations.
OpenAI’s shared prompt examples also point to broader layout control. The model was shown handling manga-style pages, multi-panel storytelling, vertical page formats, and tightly specified placement of visual elements. That pushes it closer to producing coordinated visual sets instead of just standalone posters.
Chinese poster tests still produced unreadable text
The gains did not carry over well to non-English scripts. In one media test, ChatGPT was asked to imitate a Chinese fan and create a support poster for actor Timothée Chalamet. The output was visually dramatic and packed with details including traditional clothing, cat ears, bubble tea, and pandas. It also contained more than 20 text segments, but most of them could not be read.
When the tester asked what those characters meant, ChatGPT answered bluntly that most of the text was fake, describing it as “half-gibberish AI text disguised as a Chinese meme poster.” It added that some parts were visibly distorted and mixed with characters that looked Japanese, saying the output was meant to imitate the feeling of East Asian fan-edit typography rather than form accurate sentences.
Based on the material released so far, ChatGPT Images 2.0 makes a visible leap in English text generation and expands image workflows with multi-image output, custom dimensions, and more structured prompting. Chinese and other non-English text rendering, though, remains an unresolved weakness.