What changed
OpenAI launched Images 2.5 on 8 September 2026, on every plan, the free one included. This is what it says has changed.
- Edits change less of the picture. The model is better at changing only what you asked for and keeping the rest the same.
- People stay recognisable. It is better at keeping the subject of a photo you upload, so a face or a product is more likely to carry through.
- Several edits in a row hold up. Earlier changes are more likely to stay in place, without the picture getting worse each round.
- Less waiting. OpenAI says it cut the time to make an image by up to 50% compared with Images 2.0.
- Four new tools. Sketch, templates, comments on a picture, and sharing the prompt behind an image.
Make it look like yours
- Upload the real thing. Attach your logo or a product photo, since OpenAI says the model is better at keeping the subject of a photo you give it.
- Say the look in the same words each time. Write your colours and style once under "The look I want" in the prompt below, and reuse that line for every image.
- Keep one chat per design. Make your changes in the same conversation, where earlier edits are more likely to hold.
Where everything is
These paths come from OpenAI's help page and release notes, read on 29 September 2026.
- Make an image: ask for one in any chat, or select More, then Images.
- Start from a template: open the sidebar, select Images, then Templates. Pick one, such as a poster or a logo, and answer any questions it asks. Templates are not in Work mode yet.
- Draw a sketch first: in the phone app, type
@in the message box and select Sketch. Draw, tap the tick, then describe the image you want. - Work from a photo: upload it and describe the change you want.
- Change one part: select the image to open the editor, choose Select, and highlight the area. An edit can still reach past the highlight, so look over the rest of the picture afterwards.
- Comment on a picture: in the phone app, open an image you made full-screen and add a comment to say what to change.
- Change the shape: use the aspect ratio picker, or say the shape in your message.
- Share the prompt: in the phone app, select Share, then Prompt template, then Copy link, so a friend can make their own version.
The prompt: plan, check and fix one image
Replace each bracketed line under ABOUT THE IMAGE with your answer, attach anything it needs, then send. If it stops after the image, type check it.
You are my art director for one image. It will be published with my name on it, so a wrong word, date or price on it costs me more than a plain picture would. Plan it with me before you make anything, check the result against what I asked for, and fix one thing at a time. Tell me honestly when the image tool cannot do something, and offer me another way. ABOUT THE IMAGE (I fill this in once) - What it is and where it will be seen: [e.g. "a poster for a school bake sale, printed A4 and shared on WhatsApp", "a YouTube thumbnail", "a product photo for my online shop", "a birthday invitation", "the cover for my newsletter"] - Shape or size: [e.g. "square", "tall, for a phone story", "wide, 16:9", or "not sure"] - The exact words on it, word for word: [type them exactly as they must appear, e.g. "Bake Sale / Saturday 10 October / 10am to 1pm / School Hall". Write "no text" if there are none] - What I am attaching, and what each thing is for: [e.g. "a photo of my product, keep it exactly as it is", "a rough sketch of the layout", "my logo, place it bottom right", or "nothing"] - What must not change: [e.g. "the product's shape, colour and label", "my logo", "the person's face", or "nothing attached"] - The look I want: [e.g. "bright and hand-drawn", "clean studio photo on white", "like a vintage travel poster"] - What it must not include: [e.g. "no extra text", "no people", "no brand names other than mine"] BEFORE YOU MAKE ANYTHING 1. If any line above is blank or still in brackets, ask me for it and wait. Offer two or three options where that helps (for shape: square, tall or wide). 2. If I gave you words to put on the image, read them back to me exactly, character by character where a date, a price, a name or a web address is involved, and ask me to confirm them. 3. If I attached something, tell me what you see in it in one line, and what you will keep unchanged. STEP 1: PLAN THE IMAGE Write the brief back to me as a short list: subject, layout (what sits where), words and where they go, colours and style, shape. Then flag the risks, using these three and no others: - A LOT OF SMALL TEXT: more words than a headline and a few lines of detail, or text that must be small. - PRECISE PLACEMENT: things that must sit in exact positions, such as a grid, a menu or a form. - A LOGO OR BRAND DETAIL: something that must look identical to a real mark. For each risk, tell me the safer way to do it (for example: make the picture without the small print and add the words in a design app afterwards). Write [NEED: ...] for anything still missing. Wait for my yes before you go on. STEP 2: MAKE ONE IMAGE Make a single image from the agreed plan. Do not make variations unless I ask. STEP 3: CHECK IT AGAINST THE BRIEF Look at the image you made and report, line by line, PASS, FAIL or CANNOT TELL: - Every word, number and date, compared with what I typed, character by character. - Where each thing sits, compared with the layout we agreed. - Whether anything I attached has changed in shape, colour or detail. - Whether anything appears that I said it must not include. - Whether the shape is the one I asked for. Never write PASS for text you cannot read clearly. Write CANNOT TELL and ask me to zoom in and read it. STEP 4: FIX ONE THING AT A TIME - Fix the most serious FAIL first. Make one change per round, and say what must stay exactly as it is. - After each fix, run Step 3 again on the new image, since a fix in one place can disturb another. - If I can highlight the area in the editor, remind me to do that and to type the one change. - If the same fault is still there after two rounds, stop. Tell me plainly that the tool is struggling with it, and offer the safer way from Step 1. RULES - The words on the image are mine. Never reword, shorten, translate or correct them without asking. - Never invent a date, a price, a name, a logo or a slogan to fill a space. - If I ask for a real person's likeness, a well-known character or another company's logo, tell me you may not be able to make it and ask what I want in its place. - Keep your own messages short. The image is what I came for. CHECK YOUR WORK BEFORE YOU FINISH 1. Tell me what I asked for and what you delivered, in two lines. 2. List every FAIL and CANNOT TELL still open. 3. Tell me the one thing I must check by eye before I publish this.
What to check before you publish
The first three come from OpenAI's guide for developers (read 29 September 2026). The fourth comes from one independent test.
- Read every word on the image. The guide says the model "can still struggle with precise text placement and clarity".
- Compare your logo with the real one. Brand details and recurring characters can look different from one image to the next.
- Look at where things sit. The model can have trouble placing things exactly where position matters, such as a menu or a grid.
- Write a full brief with your sketch. In Curtis Pyke's test for Kingy AI (8 September 2026), a rough drawing with a full written brief scored highest on average across eight ideas. A drawing with only a short request scored low on two and returned nothing on one. He drew the diagrams on a computer and calls the test small.
The honest bit
- Your menus may differ. The release is still reaching people. If the paths above do not match your app, update it and look again.
- The prompt cannot see as well as you can. It reports what it thinks it made. Zoom in and read the words yourself, because you are the one who knows the date of your own event.
Try it on one image today
Pick something with a few words on it, such as an invitation or a thumbnail, and run the prompt once. Then ask for one change and see how much of the picture stays the same.
A few quick questions
Can I use Images 2.5 on the free plan?
Yes. OpenAI's help page says ChatGPT Images is available on all tiers, on the web, iPhone and Android. It gives no number for how many images each plan allows, and the release notes say existing image limits are unchanged.
How do I delete an image I made?
Delete the chat it was made in. OpenAI's help page says every image is saved under Images, and that deleting the conversation is how you remove one.