OpenAI's GPT Image 2

An image model is software that turns a written prompt into pixels, and one of its oldest weak spots has been letters that come out warped or unreadable. OpenAI says its new ChatGPT Images 2.0 model is built to handle text, layouts, and edits more reliably than earlier versions. (openai.com) (developers.openai.com) OpenAI surfaced “Introducing ChatGPT Images 2.0” on its homepage on April 22, 2026, and its developer docs now list `gpt-image-2` among the GPT Image models available through the image generation tool. The company’s current ChatGPT image model page also says the default `chatgpt-image-latest` snapshot is the image model used inside ChatGPT. (openai.com) (developers.openai.com 1) (developers.openai.com 2) The API docs describe two ways to use these systems: a direct Image API for one-off generations and edits, and the Responses API for conversational, multi-turn image work. In the Responses flow, a mainline model such as GPT-5.4 can call the image-generation tool, revise the prompt, and either create a new image or edit one already in the conversation. (developers.openai.com 1) (developers.openai.com 2) That setup targets a practical problem for designers and marketers: posters, slides, menus, diagrams, and interface mockups often fail when the model cannot keep spelling, spacing, and layout straight. OpenAI’s December 16, 2025 image update already framed the product as a tool for “precise edits” and detail preservation, and the new `gpt-image-2` listing shows the company is still pushing image generation deeper into editable workflows rather than single-shot art prompts. (openai.com) (developers.openai.com) The timing also lands in a crowded image-generation market where model makers are competing on usefulness, not just visual style. OpenAI’s current model catalog still labels GPT Image 1.5 as its “state-of-the-art image generation model,” while `chatgpt-image-latest` is the default model in ChatGPT and GPT Image 1 is listed as the previous image model. (developers.openai.com 1) (developers.openai.com 2) (developers.openai.com 3) The technical shift is toward multimodal systems, meaning one model can read text, inspect images, and produce new images in the same loop. OpenAI’s docs say GPT Image models accept text prompts, can take image inputs, and can return edited or newly generated images, which is the kind of pipeline used in creative software and some visual-perception applications. (developers.openai.com) (developers.openai.com) OpenAI’s current pricing pages also show the company is positioning newer image models as cheaper to run than the older GPT Image 1 in several common settings. The `chatgpt-image-latest` page lists high-quality 1024x1024 image generation at $0.133 per image, versus $0.167 for GPT Image 1 at the same size and quality. (developers.openai.com) (developers.openai.com) Some limits remain in the official docs. The image-generation tool page says `gpt-image-2` supports flexible sizes within its resolution constraints, but it does not currently support transparent backgrounds in that tool flow. (developers.openai.com) The immediate test for ChatGPT Images 2.0 will be simple to spot: whether signs, labels, maps, and multilingual graphics now come back usable on the first try. OpenAI has moved image generation from novelty art toward document-like visual work, and the next few weeks will show how often creators trust it with jobs that used to require a human designer polishing every line of text. (developers.openai.com) (openai.com)