GPT Image 2.5, reviewed

On 8 September OpenAI released ChatGPT Images 2.5, five months after Images 2.0. In ChatGPT it became the image model for every user. In the API it arrived as two models: GPT-Image-2.5 Flare, the fast default, and GPT-Image-2.5 Sunburst, which takes longer for more precise work.

OpenAI says people now create more than 3 billion images a week with its tools. Here is what changed, what the model does well, where it still slips, what it costs, and what it means for a brand that wants to use it.

What changed

The headline claims are sharper detail, more natural light and texture, and faster results, with up to 50 percent less waiting than Images 2.0. OpenAI also says the model keeps the subject of a reference photo more recognisable and follows editing instructions more reliably across several turns, so earlier changes no longer fall apart.

ChatGPT gained tools around the model. Sketch turns a rough drawing into a finished picture. Templates start a poster or a product photo from a structure instead of a blank prompt, and comments can be placed directly on an image to direct an edit.

Much of the groundwork came in April with GPT Image 2. TechCrunch found it “surprisingly good” at text, with a restaurant menu plausible enough to use, and OpenAI pointed to better rendering of scripts such as Japanese, Korean, Hindi and Bengali. It also added a thinking step: the model can search the web, make several images from one prompt and check its own output.

Quality, text and editing

Blind votes put it on top. On Arena's text to image leaderboard, where people compare two anonymous images and pick the better one, Sunburst ranked first and Flare second on 7 October, ahead of GPT Image 2 and clearly ahead of Grok Imagine, Microsoft's MAI-Image-2.6 and Google's Nano Banana 2.1. Votes like these measure what people prefer at a glance, not whether an image is right for a brand, but the lead is consistent.

Text is where the progress shows most. Headlines, menus, labels and interface elements that came out garbled two years ago are now usually spelled correctly. OpenAI's own documentation is more careful than its launch post: the model “can still struggle with precise text placement and clarity”. In our view that means yes to drafts and mockups, and yes to final packaging or legal copy only after a person has checked every word.

For a brand, editing matters more than the first image. OpenAI says Images 2.5 is better at changing only what was asked, so a product, a background or a line of copy can be swapped while the composition stays. Its documentation names the limits plainly. A mask is guidance, not a stencil, and the model “may not follow its exact shape with complete precision”. It may also struggle to keep recurring characters or brand elements consistent across generations, and to place elements precisely in layouts.

The practical rule follows from that. Treat every edit as a new rendering of the whole picture and compare it with the original before it goes anywhere. Where a product must stay exactly as it is, such as a bottle with a regulated label, keep the real photograph and let the model build the scene around it instead of redrawing the object.

Speed, price and limits

Flare is the model OpenAI recommends for most uses. Sunburst is meant for campaign work where edits must hold, and takes longer. Both cost the same per token: 30 dollars per million image output tokens, plus input. OpenAI's own calculator puts a square image of 1'024 pixels at about half a cent in low quality, about 5 cents in high quality and about 21 cents at the new max setting.

Images can reach 3'840 pixels on the long edge, though anything above 2'560 by 1'440 is still marked experimental, and complex prompts can take up to two minutes. OpenAI may ask an organisation to verify itself before it can use the image models in the API. Models are also retired quickly: gpt-image-1 shuts down on 23 October, and GPT Image 1.5 and its mini version follow on 1 December. Every workflow built on one model needs retesting on the next.

Against Google and the rest

Google answered on 6 October with Nano Banana 2.1, built on Gemini 3.6 Flash. It takes up to 14 reference images, keeps up to four characters and ten objects consistent, renders up to 4K and costs, according to The Decoder, 3.36 US cents an image at 1K and 7.56 cents at 4K. Google's own model card lists the weak spots: small text that is often blurry at 1K, imperfect likeness to reference images and occasional confusion between left and right.

So the trade is clear. GPT Image 2.5 leads on overall preference and on text, and its editing has improved. Google is cheaper at high resolution, accepts more references at once and sits inside Google Ads, Flow and Stitch. In The Decoder's tests, Google's older and more expensive Nano Banana Pro still often produced the more natural picture. No model wins every job, which is why the choice is best made per job, not once a year.

What it means for a brand

Our verdict: on today's evidence, GPT Image 2.5 is the strongest general image model available, especially where text and layout matter. Use Flare for drafts, concepts and social formats, and Sunburst for final campaign work. Check every word, keep the real product photograph for anything that must be exact, and keep a second model ready for jobs where Google is cheaper or better. OpenAI marks its images with C2PA metadata and an invisible watermark, which helps with honesty.

It is still not a brand system. It does not know your light, your crop or your colours unless you show it, and it can drift between generations. That is the part Reel is built for: styles locked from real references, the best model chosen for each job with GPT Image and Gemini among them, a quality check on every result, and a person who picks what is used. A better model makes that process faster. It does not replace it.