OpenAI just dropped a new image generation model, ChatGPT Images 2.0, replacing its previous native image model. It's available now on the web and through the API.
The model comes in two flavors: Instant and Thinking. Instant is for speed—it was tested internally under the codename "duct tape" on LMArena. Thinking mode, on the other hand, takes a moment to reason before generating, which helps maintain character consistency across frames and handle multi-scene narratives. Think comic panels or sequential design.
On the spec sheet, Images 2.0 can output up to 2K resolution, with aspect ratios ranging freely from 3:1 to 1:3. You can generate up to eight images at once. Its knowledge cutoff has been pushed to December 2025, and it can pull in the latest information from the web before generating an image.
The interaction model has shifted from one-off generation to conversational editing. You can zoom in, change the composition, or swap elements within the same session, and the model retains context between edits. In a live demo, the model took an uploaded photo and suggested eight different summer outfits. Another example showed it scraping social media feedback on the previous model, then visualizing those conclusions and adding a QR code linking back to ChatGPT.
Text rendering and non-Latin scripts are another focus. OpenAI says performance for Japanese, Korean, Chinese, Hindi, and Bengali has improved significantly—a long-standing weak spot for diffusion models. This comes as the company continues a broader pivot toward enterprise and developer tools, following recent high-profile departures like that of Sora lead Bill Peebles.