What Powers Image Generation in ChatGPT
Image generation in ChatGPT is handled by the GPT Image family. As of August 2026 the current model in both ChatGPT and the API is GPT Image 2 (gpt-image-2), and the DALL-E 2 / DALL-E 3 models it replaced are no longer served.Generation is a capability of the language model itself rather than a separate drawing system called as a second step, so the context of your conversation carries straight into the picture. That is also why you can hand over a base image and ask for one part to be changed.
What the Integration Makes Easy
- Conversation to image: describe the figure you want; the model turns your intent into a prompt and generates it.
- Iterative editing: after generating, say "change this part like this" in plain language. You can also supply a base image and replace only part of it.
- Illustrations in context: rough illustrations and concept diagrams that match the tone of the article or deck you are writing.
Local editing (image-to-image) is practical: change the background, colors or text of an existing image without regenerating everything, and grow a single image through conversation.
Text Inside Images
Rendering readable text is a particular strength — character-level accuracy is reported at around 99%, covering headlines, fine print and multilingual labels (Japanese, Chinese, Korean, Arabic and others). Posters, ads, product packaging, menus, UI mockups and diagram drafts are all in scope. Detailed technical illustrations and exact numeric tables still need to be checked by eye before you publish.
Where It Fits, and Where Another Tool Fits
| Good fit for ChatGPT image generation | Consider another tool |
|---|---|
| Illustrations and concept diagrams that follow the conversation | A consistent series or fixed characters (dedicated services can be stronger) |
| Banners, posters and UI mockups that contain text | Meticulous artistry or a very specific style |
| Iterative editing and partial replacement starting from one image | Large batch generation (an API or dedicated pipeline is more efficient) |
Pricing and Access
Pricing and limits change often. Treat the following as a rough guide and confirm on OpenAI's official pricing page.
- ChatGPT free tier: the fast mode of GPT Image 2 is available to free users, with a cap per few hours; beyond it, requests fall back to a lighter model.
- ChatGPT Plus: about USD 20/month. Pro: about USD 200/month, with far looser limits.
- API (gpt-image-2): roughly USD 0.01 low quality / USD 0.05 medium / USD 0.2 high per image, varying with resolution and quality. At sustained volume the API becomes cheaper than a subscription.
Things to Watch
- Commercial use: generally permitted, but the terms are updated — check the primary source for client work.
- Real people and brands: imitating public figures, existing brands and characters is restricted.
- Provenance (C2PA): output carries C2PA content credentials and an invisible watermark, so AI origin stays traceable.
- Verify facts: numbers, proper nouns and technical figures inside an image can be wrong however convincing they look.
Summary
Create by talking, fix by talking, finish with text in the frame. Start by telling ChatGPT the image you want and refining the first result in conversation. Before you publish, check three things: the text, the numbers, and the rights.
