OpenAI launches ChatGPT Images 2.0 with text generation and thinking

OpenAI released ChatGPT Images 2.0 on April 22, 2026, available under the model ID gpt-image-2. The launch marks a departure from the diffusion-based architecture of earlier image generators. The new model incorporates what OpenAI calls “thinking capabilities” — before producing an image, the model searches the web, generates multiple candidate outputs, and verifies the result against the original prompt. The output is checked before it is delivered, rather than handed off immediately.

The most significant improvement is text accuracy. Previous image generators produced menus and signs with misspelled words; the running example in early coverage was a restaurant menu that read “enchuita” instead of “enchilada.” Images 2.0 creates printed text that is indistinguishable from a professional layout. The model also handles non-Latin scripts, including Japanese, Korean, Hindi, and Bengali, which had previously been unreliable across most image generators.

For product teams, the practical change is that generating marketing assets, UI mockups with readable text, and multi-format visuals now produces usable output on the first attempt rather than requiring manual correction or a second pass from a designer. The model renders small text, UI elements, and iconography at resolutions up to 2K. Multi-paneled formats, such as instructions or comic-style explanations, are handled in a single prompt.

Access opened to all ChatGPT and Codex users on April 22. Enhanced capabilities are available for paid subscribers. The model is also accessible through the API at usage-based pricing.

For product managers working on content-heavy products, e-commerce, or anything that involves regular generation of visual assets, this changes the cost and speed assumptions around that part of the workflow. The constraint is no longer whether the output contains readable text — it is whether you have a clear prompt.