ChatGPT images can move now. With Images 2.6, ChatGPT can make looping GIFs. The banner image for this article (the thumbnail on note) was also generated with Images 2.6. The cat illustration and the large Japanese text "GIFが作れる!" ("You can make GIFs!") both rendered cleanly in a single generation. In my experiment, an edit that changed only the color of a generated banner also kept the text and composition intact. This post covers how to make these images and how to phrase the requests, along with what I learned while actually making one.
Images 2.6 was announced by OpenAI on November 4, 2026. The announcement alone makes it hard to judge how useful the update will be in your own work. So I tried generation, editing, and GIFs while making the banner image for this article.
This article covers six things:
- What changed in Images 2.6 (based on OpenAI's announcement)
- How to ask for a GIF (animating a still image in ChatGPT)
- How to ask for generation, editing, and reference images
- What I tried this time, including what worked and what didn't
- Whether to use ChatGPT, Codex, or the API
- How to distinguish official claims from my test results
What changed in Images 2.6
According to OpenAI's announcement, Images 2.6 is rolling out in ChatGPT, Codex, and the API. In ChatGPT it is available on all plans, including the free plan, and works on desktop, mobile, and the web. The main changes are:
- Generation speed: OpenAI says it is 50% faster than before.
- GIFs: ChatGPT can now make looping GIFs. It can also animate a still image.
- Editing consistency: OpenAI says character, product, and brand details now hold up across repeated edits. Its example is changing the background while the outfit stays the same.
- Saved reference photos: you can save a photo under a name and call it up in later chats.
- Two API models: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst.
The 50% speedup and the improved consistency are both OpenAI's claims. In this article I keep official claims and my own measurements separate.
So how do you ask for these features? The next sections cover four kinds of requests.
GIFs: animate a still image in ChatGPT
GIF creation is a ChatGPT (web/app) feature. It can make looping GIFs and can also animate a still image. After creating one, you can adjust the playback speed, save it, and share it.
To animate a still image, attach the image and ask for something like this:
Using this still image as a base, make a looping GIF in which the light particles around the planet move slowly.
This is a sample request for ChatGPT, and I didn't try it in this project. The teaser below notes that I couldn't make a GIF through Codex.
Keep these points in mind:
- The feature is rolling out gradually. It isn't live for every account at once, so it may not show up on your screen yet.
- You can't edit a GIF that you've uploaded.
For images with text, specify the ratio, the text, and what to exclude
For an image with text, like a banner, your intent comes through more clearly if you specify more than the content: the aspect ratio, the exact text, and what must not appear. Four points matter when you write the request:
- Put the aspect ratio first: state the ratio that fits the use case right at the start, for example "wide banner image (aspect ratio about 1280:670)".
- Split the layout left and right: say which side the illustration goes on and which side the text goes on.
- Put the text in quotation marks: write each line of text inside quotes, along with differences in size and color (for example, large bold white versus smaller white).
- State what to leave out: a line like "No other text, no logos, no people." tells the model not to add extra text, logos, or people.
The complete prompt I used for this banner is in the paid guide below.
For edits, separate what changes from what stays
When you want to change only part of a generated image, write what changes and what stays the same as separate items: one item for the change, one for what to keep. My exact instruction for this project is in the paid guide below.
If you only say "change the color," it's unclear how much the model is allowed to change. Listing what must stay the same also gives you a checklist when you compare the results.
Reference images: save a photo with a name and reuse it with @name
According to OpenAI's help page, if you save a photo under a name, you can call it up in a later chat by writing "@name" and reuse it. This is useful when you keep using the same subject, such as a product photo or a character.
There are two steps:
- Save the photo under a name.
- In a later chat, call it up by specifying "@name".
I didn't try this feature in this project. The steps above are based on the help page.
What happened when I tried these requests
Here's what happened when I tried these requests. I used Codex this time, which can also generate and edit images with Images 2.6. The headline outcomes:
- Generation: the text rendered correctly in a single generation.
- Editing: the color-only edit kept the text and composition.
- GIF: it did not animate via Codex.
Get the complete prompts and test log ($5)
The free part covered what changed in Images 2.6, how to phrase each kind of request, and the headline results. The paid document adds the verbatim banner prompt, my exact edit instruction, and the full test log, including image sizes, run times, and a result I had to redo.
What's inside:
- The complete banner prompt, ready to copy, with the four points read line by line
- The exact edit instruction I used (what changes, what stays)
- Generation result details: output size and run time
- Edit comparison details, including what was and wasn't preserved
- The second-attempt story: why my first edit couldn't count as evidence, and how I redid it
- The GIF result via Codex
ChatGPT, Codex, or API: choose by what you want to do
The available Images 2.6 features depend on whether you use ChatGPT, Codex, or the API.
- ChatGPT: where you get the on-screen features, such as GIF creation and saving and calling up reference images.
- Codex: can generate and edit images. GIF creation is a ChatGPT feature, and I couldn't make one through Codex in this project.
- API: you can choose between two models.
- GPT-Image-2.6 Flare: the speed-oriented model. At low quality it takes as little as 6 seconds, according to OpenAI. Layout quality for things like posters and product images has also reportedly improved.
- GPT-Image-2.6 Sunburst: a model for medium/high quality, aimed at fine detail and editing consistency.
Separately from the models, the API also has new features: image markup, SVG export, and a complexity parameter. Image markup lets you mark up the areas you want to edit; it already existed in ChatGPT and has now been added to the API. SVG export lets you export an image as an SVG. The complexity parameter adjusts how much planning the model does; low is recommended for most edits, and medium or high is said to suit scenes with many elements.
API pricing is on OpenAI's pricing page. Flare and Sunburst cost the same, and that is the same as 2.5. Per 1 million tokens, image input is $8, image output is $30, and text input is $5. The cost per image depends on image size and quality, so I don't state a figure here.
Read official claims and measurements separately
Keep these three points in mind when reading this article.
First, treat official claims and measurements as separate things. The 50% speedup and the improved editing consistency are OpenAI's claims. What I confirmed in this project is limited to this: generation came out readable in one shot, and in a color-only edit the text and composition were mostly preserved by eye. I did not verify the speedup itself.
Second, some features are still rolling out. For features such as GIF creation that are rolling out gradually, check your own account to see whether they're available.
Finally, generation times weren't compared rigorously. Times vary with the route and how you run it. These approximate times cover entire Codex runs, so they can't be directly compared with OpenAI's claimed 50% speedup.
Start by deciding the text for your own image and generating just one
The smallest thing you can try today is a single generation. A banner or an image for social media will do. Put the text you want in quotation marks, add the aspect ratio and what must not appear, and give that prompt to ChatGPT or Codex. You can get started by applying the four points above to your own subject and text.
Once you have an image, try one edit as well, with the instruction split into what changes and what stays. How well the text and composition hold up is most reliably judged on your own image. But first, generate just one image.
Originally published in Japanese on note. This English edition was translated and edited with AI assistance; all commands and measurements are unchanged from the original.
Top comments (0)