Generate and edit images online with OpenAI's newest model. Write a prompt, add reference photos, and change one detail at a time while the rest of the frame holds.
Name what has to stay the same in every edit: keep the face and the lighting, change only the jacket.
GPT Image 2.5 is the image model OpenAI released on September 8, 2026, and it ships inside ChatGPT as ChatGPT Images 2.5. Developers call it gpt-image-2.5. It reads text and pictures together, so a prompt, a reference photo, and a follow-up correction all belong to one conversation. Set against ChatGPT Images 2.0 from April 2026, it renders finer detail, answers up to 50% faster, and keeps an edit chain from falling apart.
Change one element and the subject, composition, background, and brand treatment you never mentioned stay exactly as they were.
Latency drops by up to half against ChatGPT Images 2.0, and Flare runs 2 to 4 times faster than GPT Image 2 at matching quality.
Posters, packaging, UI mockups, and infographics come back with legible small print instead of scrambled letterforms.
OpenAI ships gpt-image-2.5-flare for everyday speed and gpt-image-2.5-sunburst when editing precision matters most.
From an empty prompt box to a finished, edited image, with nothing to install.
Write your subject, composition, style, and lighting in the prompt box, plus any words that must appear in the picture, quoted exactly as you want them printed.
Upload a face, a product, or a style board when identity matters, then say what each one is for: image 1 is the subject, image 2 sets the palette.
Pick an aspect ratio and an output size, up to 4K on the long edge, then generate and check the result against your brief.
Send a follow-up naming the single change and restating what must stay identical, repeat until the frame is right, then download it.
Four capabilities that decide whether an image model survives production work, all live in the generator above.
Designers, marketers, and builders on their first week with GPT Image 2.5.
We rebuilt a campaign key visual in one sitting. Eleven edits deep, the model's face and our packaging still matched the originals, which never happened before.
Sofia Marchetti
Brand Creative Director
Transparent backgrounds are why I stayed. I generated 60 item icons and HUD pieces, dropped them into the engine, and only two needed cleanup. That used to cost me a weekend.
Kenji Watanabe
Indie Game Artist
Flare is quick enough that I stopped batching work overnight. We ship around 30 post variants a day, and the 80s flashback prompts alone doubled saves on our last three posts.
Priya Raman
Social Content Lead
I prototyped an in-app generator on gpt-image-2.5-flare in an afternoon, then moved the launch page heroes to sunburst. Same prompts, two speed profiles, no rewriting.
Owen Whitaker
Product Engineer
Background swaps used to wreck the product. Now I say keep the bottle and the label, change the shelf, and that is what comes back. We refreshed 140 listing images in two days.
Nia Bankole
Ecommerce Visual Merchandiser
Typography was the dealbreaker before. I asked for a quoted headline and a date line in small print, and both came back readable at full size.
Lucas Ferreira
Poster Designer
What people ask before running their first GPT Image 2.5 prompt.
Write one prompt, add a reference, and let GPT Image 2.5 handle the fine detail, the readable type, and the surgical edit.
