On 8 September 2026 OpenAI released ChatGPT Images 2.5, the new version of the image generator built into ChatGPT. According to the company it is rolling out to every plan, the free one included, on desktop, mobile and web. Generation is up to 50% faster than Images 2.0. The model is better at keeping the subject of a reference photo intact instead of redrawing it, and it follows editing instructions more reliably across several rounds.
For a small business this touches one specific job: the routine visual material. An Instagram post, a flyer for a sale, a product photo for the online shop, a banner for the business page. Most sole traders either pay a designer for that, buy a ready template, or skip it.
Four new tools in the interface
The first is called Sketch. You type @Sketch in the chat, draw a rough layout with a finger, mouse or stylus, and add a written instruction. The model uses the drawing as a reference and produces a finished image from it. Someone who knows where the logo should sit and where the price goes, but cannot design, gets a way to show it instead of describing it.
The second is templates: ready starting points for common formats, including posters, merchandise, flyers and product photography. A template spares you the blank page and the work of writing a long instruction from scratch.
The third is comments on the image. You mark one area and ask to change only that, and the rest stays where it was. That addresses the familiar frustration with these tools, where a small fix to a headline produces an entirely different picture.
The fourth is prompt sharing. A shared image can carry the prompt that made it, so a partner or an employee can produce another version in the same style without guessing what was written.
Where it replaces a designer, and where it does not
Product photos are where the improvement shows most. You upload a phone shot of the product and ask for a clean background, different lighting or a whole scene around it. Earlier versions tended to redraw the product, sometimes changing its shape, colour or label. This one was built to keep it as it is. The result is still worth comparing with the original photo, because a customer who unboxes something different from what the picture showed complains to the business, not to the tool.
Text inside the image remains the weak part. In this version too, lettering in the picture still slips at times. Anyone putting a price, a sale date or a phone number into a post needs to read every character after generation, and in Hebrew to check the direction of the line as well. The safer route is to generate the image with no text at all and add the words afterwards in an ordinary editor where you can see exactly what was typed.
A picture for this week's post and a brand identity are two different jobs. A designer who builds a consistent visual language for a business, with colours, fonts and rules that repeat across every piece, does something a model producing one image on request does not. Prompt sharing brings some consistency closer, but it does not replace a settled decision about how the business looks.
The price on the API and on a plan
On a ChatGPT plan the tool is included in the subscription you already have. The announcement does not list generation limits per plan, so anyone on the free account will find the ceiling when they hit it.
Anyone wiring image generation into an online shop or an automation works with two API models. Flare is the fast default. Sunburst is slower and meant for precise editing. Both carry the same price: 5 dollars per million text input tokens, 8 dollars per million image input tokens, and 30 dollars per million image output tokens. That is the same per-token price as the previous generation, so anyone already running such a pipeline gets the improvement with no change to the rate.
Three minutes before the post goes up
An ad with a wrong price or a product that looks different goes out under the business's name, even when a model made it. A routine that holds up: generate from a real photo of the product, compare the result with the original, add the text by hand, and keep the prompt that worked so the next post looks like the last one. It takes a few minutes, and the alternative is a public correction in the comments.