What is GPT-4o image generation in ChatGPT, and what can it do?

Yesterday, Tuesday, March 25, 2025, OpenAI changed how ChatGPT makes pictures. Until now, when you asked for an image, ChatGPT passed your request to a separate model, DALL-E 3. From yesterday, GPT-4o, the model you already chat with, draws the image itself. OpenAI calls this 4o image generation.

That sounds like a technical detail, but you notice the difference quickly. Signs, menus and posters come out with readable words. Long, specific requests are followed more closely. And you can keep asking for changes in plain language while the picture stays consistent.

What OpenAI released

OpenAI says the feature rolls out starting yesterday to Free, Plus, Pro and Team users as ChatGPT's default image generator, with Enterprise and Edu users to follow. TechCrunch reports that Pro subscribers, on the $200-a-month plan, got it first, with the others close behind. It also works in Sora, OpenAI's video app. Developers will get access through the API in the coming weeks. If you prefer the old model, DALL-E remains available through a dedicated GPT, one of the custom versions of ChatGPT.

There is a trade-off: the images are more detailed and take longer to appear, often up to a minute.

How "native" image generation works

OpenAI describes GPT-4o as natively multimodal: one model that works with words and pictures alike. Being native means image creation is one of the model's own skills, not a job it hands to another program.

An analogy: with the old setup, ChatGPT was like a client describing a picture to an illustrator in another building. Some details got lost in the message. Now the person you are talking to holds the pencil. It knows everything said in the conversation, it can look at a photo you upload, and it can draw on its general knowledge, for example what a correct recipe card or a weather chart should contain.

OpenAI says it trained the model on images and text together, so it learned how pictures relate to words and to each other, and then refined it heavily afterward. TechCrunch adds that OpenAI told the Wall Street Journal the training data includes publicly available material and licensed data from partners such as Shutterstock.

Four strengths stand out in OpenAI's examples:

• Text in images. Street signs, labels and menus with the words spelled as asked.
• Many objects at once. OpenAI says GPT-4o can keep 10 to 20 distinct objects straight in a scene, where other systems manage about five to eight.
• Editing by conversation. You can ask for a wider format, a new background or a different style, and characters keep their look from one version to the next.
• Working from your images. Upload a photo or sketch and ask the model to transform it or use it as a reference.

You can also state practical details: the aspect ratio (the shape, such as wide 16:9), exact colors as hex codes like #1A73E8, or a transparent background for a sticker or logo.

Limits and safeguards

OpenAI lists known weaknesses. The model sometimes crops tall images such as posters too tightly at the bottom. Like any language model, it can invent details, especially when the prompt says little. It struggles with very small text, with writing in non-Latin alphabets, and with very dense subjects such as a full periodic table. Asking it to fix one small part, such as a typo, may change other parts of the image too.

Every image carries C2PA metadata, a standard digital label that records the picture came from GPT-4o. OpenAI says it keeps blocking harmful content, including sexual deepfakes, and applies stricter rules when real people appear in an image. Its chief operating officer, Brad Lightcap, told the Journal that OpenAI has policies meant to stop it from directly imitating the work of living artists.

Why it matters

For most people, image generators have been good at striking, imaginative scenes and poor at everyday visuals: an invitation with the right date, a labeled diagram, a menu with prices. OpenAI's stated aim is to make images useful for communication, not just decoration, and the examples it released lean heavily on exactly those everyday jobs.

It also raises the bar among competitors. Google recently tested native image output in Gemini 2.0 Flash, which drew criticism for weak safeguards. Expect image generation to become a standard skill of chat assistants rather than a separate tool.

How do you write a good image prompt for GPT-4o?

Write it the way you would brief a designer. Say what the image is for, what must appear in it, and the exact words to print, in quotation marks. Describe the style and the layout, and name the format if it matters. Because GPT-4o reads the whole conversation, you do not need to cram everything into the first message; start with the essentials and refine.

Here is an example:

Design a poster for a community bake sale. Square format, cheerful hand-drawn style,
pale yellow background. Show three cakes on a wooden table.
Print exactly this text, large and easy to read:
"Saturday Bake Sale" at the top, "May 3, 9 a.m. to noon" in the middle,
and "Community Hall, 12 Main Street" at the bottom.

Then check the result like a proofreader. Read every word, count the objects you asked for, and look at the edges for cropping. If something is wrong, name the exact fix ("change 'noon' to '1 p.m.' and keep everything else the same") rather than starting over. For anything you will print or publish, give it one last careful look yourself.