GPT Image 2

GPT Image 2 is the image model from OpenAI that powers ChatGPT Images 2.0. It came out on 21 April 2026, arriving the same day in ChatGPT, in Codex and for developers through OpenAI's API. It can make new pictures from a prompt and edit existing ones, and it is especially good at text: menus, posters and comic strips come out with lettering that reads properly, including in Japanese, Korean, Hindi and Bengali. It can make images up to 2K in size.

Its big new trick is a thinking mode. Before drawing, the model can plan the layout, search the web for facts and check its own work, and it can produce a matching set of images from one request. Everyone using ChatGPT got the basic version, while the thinking features were kept for paying subscribers. Complicated jobs, like a multi-panel comic, can take a few minutes rather than seconds.

Who makes it

OpenAI is the San Francisco company behind ChatGPT. It was one of the first to show the public what AI image generation could do, with the DALL-E models, and its image tools now reach a vast audience because they are built straight into ChatGPT.

History

DALL-E, DALL-E 2 and DALL-E 3 came first. In March 2025 OpenAI replaced them in ChatGPT with image generation built on GPT-4o, which worked quite differently and was far better at following instructions. It was an instant hit: OpenAI said people made more than 700 million images in the first week, and the internet filled with photos redrawn in the style of Studio Ghibli. That trend also drew criticism, given the studio co-founder Hayao Miyazaki's well-known dislike of AI art.

The same model reached developers as gpt-image-1 in April 2025, followed by a cheaper mini version that October and GPT Image 1.5 in December 2025. GPT Image 2 was the next big step, and OpenAI has since followed it with GPT Image 2.5.

Telling it apart

OpenAI says images made with ChatGPT, Codex and its API carry two hidden signals: C2PA metadata, a kind of digital label recording where the file came from, and an invisible SynthID watermark. You can check a picture at openai.com/verify. But metadata is easily stripped when a picture is saved, edited or shared, so a clean result does not prove an image is real.