OpenAI shipped ChatGPT Images 2.5 yesterday, and the headline feature is not what you think. Yes, the images are sharper. Yes, it is faster. But the real upgrade is what happens after the first picture appears.
You can now draw a rough sketch inside ChatGPT and have it turn into a finished image. You can place comments directly on specific parts of an image and tell ChatGPT to change just that element. And the model actually remembers what you already approved across multiple edits, so your third revision does not destroy what your first revision got right.
That last part is the big one. Every creative who has used AI image generation has hit the same wall: you get a result that is 90% there, you ask for one tweak, and the model rewrites the whole thing. ChatGPT Images 2.5 is built to stop doing that.
Here is what landed, what it can do, and how to use it right now.
What Actually Changed
ChatGPT Images 2.5 is both a model update and a product update. The model itself produces images with more natural lighting, richer textures, and better preservation of subjects from reference photos. Editing instructions are followed more precisely, especially across multiple turns.
The product side adds four new tools inside ChatGPT:
Sketch. Type @Sketch in your chat box and a drawing canvas opens. Draw whatever you want, add a text description, and ChatGPT turns your rough sketch into a finished image. Stick figures, layout mockups, outfit contours, room layouts. It all works. You do not need to be able to draw well. The sketch is a visual guide, not a finished piece.
Image comments. You can now place a comment on a specific part of an image. Instead of typing “change the background” and hoping ChatGPT understands which part, you point at the section you want changed and tell it what to do. This is surgical editing, not full-image regeneration.
Templates. Instead of starting from a blank prompt, you can pick a template like Poster or Merch and fill in your details. The template gives structure to what you are making, so you spend less time describing format and more time on content.
Prompt sharing. When you share an image, you can now attach the prompt that made it. Someone else can take your exact starting point, swap in their own photos or details, and make their own version. OpenAI linked to a viral 1980s headshot prompt as an example. The sharing mechanic turns prompts into templates that spread on their own.
The Editing Problem This Actually Solves
The single most expensive failure in AI image generation is not bad first outputs. It is what happens on the second, third, and fourth round of edits.
You generate a product shot. The lighting is perfect. The product looks right. But the shadow falls slightly wrong. You ask to fix the shadow. The model fixes the shadow and also changes the product colour, moves the camera angle, and regenerates the background.
That cycle, fix one thing and break three others, has been the biggest practical barrier for anyone using AI images in a real production workflow. Product photographers, social media managers, brand designers. They all hit it.
ChatGPT Images 2.5 is specifically designed to preserve what you already approved. OpenAI says subjects, textures, backgrounds, and composition hold more reliably across edits. Early community tests show a nine-edit sequence where the face, texture, and background stayed consistent throughout. One test is not a guarantee, but it points to the right problem being addressed.
For developers, the API has two variants. GPT-Image-2.5 Flare is the default, balancing quality and speed at up to 50% lower latency than Images 2.0. GPT-Image-2.5 Sunburst trades speed for tighter control across edits, aimed at production-ready campaign creative and polished product imagery.
What You Can Actually Do With This Right Now
Here are the workflows that just got meaningfully better.
Product photography with targeted edits. Generate a product shot, then use image comments to point at the background and ask for a studio setting instead. Or point at the product and ask for a colour variation. The rest of the image should hold.
Sketch-to-social-post. Use @Sketch to rough out a layout for an Instagram carousel or a poster. Add text describing the style you want. ChatGPT turns the sketch into a finished visual. This removes the blank-canvas problem entirely.
Iterative brand assets. Start with a prompt for a brand visual. Comment on specific elements to adjust. Swap colours, reposition elements, change the mood. Each edit should preserve what you already liked. This is the workflow that previously required Photoshop after generation.
Shareable prompt templates. Create a visual style that works for your brand. Share the prompt with your team. Everyone gets the same starting point and customises from there. For small teams without a dedicated designer, this is a genuinely useful shortcut.
Multi-reference product shots. The model is better at keeping subjects recognisable when you provide reference photos. Upload your product, describe the setting, and the output holds the product’s distinctive features more reliably than before.
What to Look Out For
The comment-based editing is the feature to watch. If it works as precisely as OpenAI describes, it changes the workflow from “generate and pray” to something closer to actual image editing. If comments sometimes drift or misinterpret which element you are pointing at, it will feel like a faster path to the same frustration.
Sketch is useful but it is not a technical first. Adobe Firefly already supports sketch-guided structure references and drawn markup. The difference here is that Sketch lives inside ChatGPT, which is where a lot of people already are. Convenience of access matters more than novelty of technique.
Prompt sharing is a distribution play. It keeps creative loops inside ChatGPT rather than sending people to other canvases. That is smart for OpenAI. Whether it is smart for your workflow depends on whether you want your prompts and their variations living inside a chat interface versus a proper design tool.
The pricing picture is incomplete. The API token rates for Flare and Sunburst are reportedly twice those of GPT-Image-2, so lower latency does not automatically mean lower cost. If you are running high-volume image generation, run the maths before you switch.
The model has not yet appeared on independent leaderboards, so the quality claims are OpenAI’s own for now. GPT-Image-2 still leads the Artificial Analysis text-to-image benchmark, and Microsoft MAI-Image-2.6 leads the editing benchmark as of September 9. Whether Images 2.5 overtakes either remains to be seen.
Adobe is also a launch customer. GPT-Image-2.5 models are now inside Firefly, Adobe’s creative AI studio. That is a significant distribution channel and a signal that Adobe is positioning Firefly as an interface layer over multiple models rather than only its own.
How to Try It
Images 2.5 is rolling out now to all ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web.
For Sketch: open a chat, type @Sketch, and draw on the canvas that appears. Add a text description of what you want the final image to look like.
For image comments: generate an image, then tap or click on the part you want to change. A comment box appears. Describe the change you want.
For templates: start a new image prompt and look for the template options in the format selector.
For prompt sharing: after generating an image, use the share function and choose to include the prompt. The recipient gets a link they can open in ChatGPT and modify with their own inputs.
For developers: GPT-Image-2.5 Flare and Sunburst are available through the Images API now.
Want more like this? Join MOKU Club for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
