ChatGPT Images 2.5 for Marketers: What Changed, What to Try First, and What It Still Cannot Do
On September 8, 2026, OpenAI released ChatGPT Images 2.5 and called it the company’s new state-of-the-art image model. If you make marketing images in ChatGPT, the release notes read like a list of your complaints from the last six months: the reference photo that stopped looking like your product on the third edit, the “make the headline bigger” instruction that redrew the whole scene, the forty-second wait for a social tile.
This post is the marketer’s read of the release, written on launch day. It covers the five changes that actually matter for marketing work, what to try first with each one, what the two new API models mean if you pick the tool your team uses, and the ceiling that did not move. Everything here is dated September 8, 2026; OpenAI ships changes to this product every few weeks, so check the linked sources before you build a process around any single detail.
The five changes that matter for marketing work
OpenAI’s own summary is that Images 2.5 “produces more natural lighting and richer textures, is better at preserving the subjects in your reference photos, and follows editing instructions more reliably across multiple turns,” with image generation latency cut by up to 50% compared with Images 2.0 (9to5Mac). Here is what each of those means when the image has a job.
1. Reference subjects survive the edit
This is the headline change for anyone with a product, a mascot, a founder headshot, or a brand illustration style. Earlier versions were good at starting from a reference and bad at keeping it: by the second or third edit, the bottle had a different label, the character had a different jawline, the palette had drifted warm. Images 2.5 is designed to keep a generated asset anchored to the original reference while you change the setting, style, composition, or an individual element. Simon Willison’s launch-day test, adding a character into an existing chart without disturbing the data, is a clean demonstration of the same behavior.
Try this first: Upload one real product photo and ask for it in three settings (on a desk, in a hand, on a plain color background from your palette). Then ask for one edit on each. Compare all six to the original. If the subject holds through the edit, you have a reference workflow. If it drifts on the edit step, note where, because that is the boundary you will be working inside.
2. Multi-turn edits hold
The second complaint the release addresses is instruction drift: you ask for one change and get three. Images 2.5 “follows editing instructions more reliably across multiple turns,” and the new comments feature makes the instruction itself more precise. You pin a note to a spot on the image and ChatGPT acts on that spot; each edit becomes a separate version you can go back to (felloai). For marketing images this is the difference between “regenerate until it is right” and “fix the one thing that is wrong.”
Try this first: Take an image that is 90% right. Instead of rewriting the prompt, pin a comment on the wrong element (“this shadow is too heavy”, “replace this icon with a simpler one”) and check whether the rest of the image stays put. Count how many turns it takes to reach done. That number, not the first-draft quality, is what determines how much of your afternoon this tool costs.
3. Sketch: layout by drawing, not by paragraph
Type @Sketch and you can draw directly inside ChatGPT, then pass the drawing to the model as a visual guide (Unite.AI). Anyone who has written “product on the left third, headline area top right, leave the bottom quarter empty for a button” and received a centered product with text everywhere understands why this exists. Spatial instructions are the thing language is worst at and a rough rectangle is best at.
Try this first: Sketch the layout of your standard social tile: where the subject sits, where the copy goes, where the logo lives. Generate three concepts against the same sketch. This is the fastest way to find out whether the model will respect a layout system, which is the foundation of a consistent set.
4. Templates: Poster and Merch
Templates give the model a predefined creative format so you are not starting from a blank prompt. The launch names Poster and Merch, aimed at flyers and product photos. They are guided prompts more than design templates in the Canva sense: they structure the request, they do not give you layers.
Try this first: Run your next event flyer through the Poster template and through a plain prompt, side by side. The template tends to produce a more finished composition; the plain prompt gives you more control. Decide which you want as a default for that asset type and write it down.
5. Half the wait, and prompt sharing
Up to 50% lower latency does not sound like a marketing feature until you remember that iteration is the whole job. Ten edits at twenty seconds is a coffee break; ten edits at forty is a meeting you skipped. Prompt sharing rounds this out: when you share an image, the prompt travels with it, so a teammate can rerun the concept with their own photo instead of reverse-engineering your wording.
Try this first: Share one finished image with the person who most often asks you for “one like that but for X.” If they can produce a usable variant without you, you have just removed yourself from a queue.
A note on the quality claims. Reviewers writing on launch day describe fewer noise artifacts and repeating surface patterns, more natural color with less of the “AI look,” stronger character consistency, and sharper text. Those are the right things to have improved. They are also claims from a first day of testing. Run the six tests in How to Choose an AI Image Generator for Marketing on your own brand before you change a process.
Access and the two API models
Images 2.5 is rolling out to all ChatGPT, ChatGPT Work, and Codex users across all tiers, on desktop, mobile, and web. There is no paid gate on the model itself. Generation quotas still differ by plan, and OpenAI did not publish updated per-plan limits with this release; third-party trackers put the free tier at a handful of images per day and Plus at roughly 50 per three-hour window, numbers that fluctuate with load and should be treated as approximate.
For developers, and for anyone evaluating a marketing tool that generates images for you, OpenAI shipped two API models alongside the chat release:
| Model | Positioning | When it fits |
|---|---|---|
gpt-image-2.5-flare | The default. Higher quality than GPT Image 2 at 50% lower latency. | Social tiles, blog covers, drafts, anything high-volume |
gpt-image-2.5-sunburst | Longer generation, tighter control across successive edits. OpenAI names “production-ready campaign creative and polished product imagery.” | The hero image that will go through three rounds of edits; product shots |
Both share one rate card, and per OpenAI’s model pages it is the same one GPT Image 2 uses: $5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens, with six quality settings from low to max (Flare, Sunburst). We go deeper on the choice in GPT-Image-2.5 Flare vs Sunburst. The short version: if a vendor tells you which of the two they run for which job, they have thought about your images. If they cannot answer, they have not.
The ceiling that did not move
Every one of the five changes makes the image better. None of them touches the marketing, and it is worth being precise about the boundary, because the better the image gets, the more tempting it is to run your whole visual program out of a chat window.
No brand memory between chats. Images 2.5 keeps a subject consistent inside a conversation. Open a new chat tomorrow and the model knows nothing about your palette, your illustration style, the things you never show, or the image you approved yesterday. Memory features carry preferences, not a visual system. Every consistent set you have seen from ChatGPT was held together by a person re-pasting the same instructions, which is exactly the failure mode described in Visual Brand Consistency at Scale.
No calendar, no campaign. The model does not know that this image is the header of Thursday’s send, that the same offer already has a LinkedIn tile from last week, or that the launch moved. It makes an image. Where the image goes, and whether it should exist at all, is entirely on you.
A flat file at the end. What you get is a PNG with an inconsistent filename (a long-running complaint on OpenAI’s own forum). Resizing for four channels, placing it in the email, the post, and the page, and keeping the original for later edits are all manual. Comments make edits cheaper; they do not make the file any less flat.
No approval trail. The image that ships is whichever one somebody downloaded. There is no record of which prompt, which references, and which version was approved, which matters the first time a customer, a lawyer, or your own future self asks where an image came from.
None of this is a criticism of the product. ChatGPT is the best place there is to try an image idea, and 2.5 makes it better at that. It is a criticism of the workflow that grows around it by default: draft in ChatGPT, download, rename, open Canva, resize, download again, upload to the ESP, repeat for social. We counted that tax in The Hidden Cost of Copy-Paste Marketing, and Images 2.5 does not lower it.
A workflow that survives the ceiling
If ChatGPT is your image tool this quarter, three habits capture most of the 2.5 upside without letting the ceiling cost you:
- Keep one reference set and one instruction block, and paste both into every image chat. Three to five reference images (your product, one photo-style sample, one illustrated sample, a palette swatch) plus a short list of rules and exclusions. The model’s improved reference fidelity is only useful if the references are always present. Turning Brand Guidelines into an AI Image Brief is the template.
- Edit with comments, not with new prompts. Reprompting re-rolls everything. Comments fix one thing. Your accepted-image rate goes up and your cost per accepted image goes down.
- Name and file the approved version the moment it is approved. Slug, channel, date, in one folder. This sounds trivial and is the single most common cause of “which one did we use?” a month later.
Where Marqeable fits
Marqeable generates images with OpenAI’s GPT Image 2.5 models inside the piece the image belongs to, so the prompt already knows the email’s subject line, the post’s copy, and the palette, imagery style, and exclusions from your brand profile, and the approved image lands in the draft with its prompt and references recorded. The reference fidelity and edit stability above are exactly why we moved to 2.5; the brand memory, the calendar, and the approval step are what we add around it. If you would rather see it on your own brand than read about it, we are in private beta with a small early cohort: get early access.
Frequently asked questions
What is new in ChatGPT Images 2.5?
Better preservation of subjects from reference photos, editing instructions that hold across multiple turns, more natural lighting and texture, and up to 50% lower latency than Images 2.0. The chat product added Sketch (@Sketch), on-image comments, Poster and Merch templates, and prompt sharing. Two API models, Flare and Sunburst, shipped the same day.
Do I need a paid plan?
No. The rollout covers all ChatGPT, ChatGPT Work, and Codex tiers. Quotas still differ by plan and were not updated in the announcement.
Should my team switch from Midjourney or Canva to ChatGPT for images?
Not on the strength of a launch post. Run the same ten prompts through each tool with your own references and score text accuracy, reference fidelity, consistency across the set, and how many edits it takes to reach done. Then decide. Midjourney Alternatives for Marketers and ChatGPT vs Canva for Marketing Images cover the trade-offs in detail.
Does Images 2.5 fix text in images?
Launch-day reviewers report sharper text, and GPT Image was already the strongest mainstream model for readable words. Whether the headline should be baked into the image at all is a separate question; Text in AI Images After GPT Image 2.5 is the decision guide.
The bottom line
ChatGPT Images 2.5 is a real upgrade for the part of the job that happens inside one image: the reference holds, the edit lands where you pointed, the layout can be drawn instead of described, and the wait is shorter. Try the reference test and the comment test today; they will tell you more than any review. Then be honest about the boundary. The model still does not know your brand tomorrow, your calendar ever, or where the file goes next. Better images do not shrink that gap. A workflow does.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
