ChatGPT Images 2.5 is here. What it means for your business

On 8 September 2026, OpenAI announced ChatGPT Images 2.5, its new image generation model. In plain terms: the pictures ChatGPT makes for you are sharper, they arrive faster, and when you ask for one change you get one change instead of a redrawn picture.

The part most coverage skips sits in OpenAI’s own release notes rather than the announcement. Two of the three new features are on mobile first, the templates are missing from Work mode, and the number of images you are allowed to generate is exactly what it was yesterday. This is a quality upgrade, not a bigger allowance.

What actually changed?

Images 2.5 replaces Images 2.0 as the model behind image generation in ChatGPT. OpenAI says it gives more natural lighting and richer textures, holds on to the subject of a reference photo more reliably, and follows editing instructions across several rounds without the picture degrading. It also cut image generation latency, the wait between asking and seeing, by up to 50% compared with Images 2.0.

Three things are new in the app itself. Templates give you a starting point for common formats, with Poster and Merch named as examples, and flyers and product photos called out in the announcement. Sketch lets you draw a rough shape in ChatGPT and have the model build the finished image around it. And you can put a comment directly on part of an image to say what should change there, rather than describing the spot in words.

For a sense of scale, OpenAI says more than 3 billion images are created every week across ChatGPT Images and its GPT-Image models.

Source: OpenAI, Introducing ChatGPT Images 2.5, 8 September 2026. Rollout detail from the ChatGPT release notes, same date.

Does it cost extra?

No. OpenAI says Images 2.5 is rolling out to ChatGPT, ChatGPT Work and Codex users across all tiers, on desktop, mobile and web. A free ChatGPT account gets the same image model as a paid one, which is not something you can say about most of this year’s launches, including GPT-6 Astra five days earlier.

There are two paid versions for developers in OpenAI’s API, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. If nobody at your company writes code, you can ignore both.

Do I get more images now?

No, and this is the line worth reading twice. OpenAI’s release notes put it plainly: “Existing image-generation limits are unchanged.” The pictures got better. The number of them you can make did not. If you were running out of generations halfway through a busy Friday, you still will.

Where do I find the templates and Sketch?

In ChatGPT, open Images and then Templates, choose one, and fill in the details; ChatGPT may ask follow-up questions to tighten the brief. Sketch is a mobile step: type @ in the message box, pick Sketch, draw your idea, then describe what it should become. Editing an image and adding comments to it also happens on mobile for now, by opening a generated image full screen.

One gap to plan around. The release notes say templates are not yet available in Work mode, so if your team sits on ChatGPT Work, the template shortcut has not reached you.

Is the editing really more precise?

That is the claim doing the most work, and the most useful early evidence comes from a company building on the API version. Axultan Alimkulov, Head of Product at Higgsfield AI, said of GPT-Image-2.5 Flare: “What impressed us most about GPT-Image-2.5 Flare is how well it understands what not to change.”

Treat that as a partner’s view of the developer model, not an independent test of the ChatGPT app, and check it on your own photos. The test takes a minute: make an image, ask for one small change, and see whether everything else survives.

Can anyone tell an image came from Ai?

In principle, yes. OpenAI says it continues to use C2PA metadata and invisible watermarking to help identify images made with its tools. C2PA is an industry standard that tucks a record of how an image was made into the file itself.

Do not rely on it as proof, though. That metadata is often stripped when an image is uploaded to a social platform or pushed through a messaging app. And whatever made the picture, the claim it makes still has to be true: a generated product shot showing a feature your product does not have is a misleading advert, and no watermark fixes that.

What to ignore

The Snake game. OpenAI’s release notes confirm that some users will see animated dots and a game of Snake while an image generates, and that is most of what the internet noticed on launch day. It changes nothing about what you get.

Ignore the pressure to switch tools over this, too. xAI’s Grok, Google and Meta have all shipped image generators this year, and for the job of making a leaflet the gap between them is far smaller than the gap between a clear brief and a lazy one. Start in the tool you already have an account for, including the design software you already pay for.

What should a small business do about ChatGPT Images 2.5?

Two steps, both free. First, open ChatGPT on your phone, go to Images, choose the template closest to what you need, and make one real thing you would otherwise have paid for or gone without: a price list card, a market stall sign, a post for an offer starting next week.

Second, photograph your best-selling product three times on a plain surface in daylight, upload the shots as references, and ask for one clean version. Judge Images 2.5 on how much of the real product survives, not on how pretty the result looks. If the answer is “close, but not quite”, that is usually a brief problem rather than a model problem, and it is the kind of thing one Breezi membership hour sorts for good.