The first build, and the receipt that comes with every image
GPT Image 2.5 shipped on 8 September with two API models, Flare and Sunburst, that ChatGPT will not let you pick between. This site prints which one answered.
OpenAI released ChatGPT Images 2.5 on 8 September 2026, along with two API
models: gpt-image-2.5-flare and gpt-image-2.5-sunburst. This is the first
build of a site that calls them.
Two models, and only the API can pick
The developer documentation is more specific than the announcement, and the difference matters:
GPT Image 2.5 Flare is the small model, optimized for speed, with image quality comparable to GPT Image 2. GPT Image 2.5 Sunburst is the base model, optimized for quality, with higher image quality than GPT Image 2. Both models offer improvements in precise editing and subject preservation.
So "2.5 is better than 2" is only true of Sunburst. Flare's value is the same picture at half the latency. Both improve on editing and on keeping a subject recognisable, and that part is the real generational change.
Inside ChatGPT and Codex you cannot say which of the two you want — the image tool does not expose the choice, and does not report which model answered. Here it is a toggle, and the answer is printed under the result.
Every image carries a receipt
Under each generation this site prints the model ID, the quality tier, the pixel size, the output format, whether the PNG actually carries an alpha channel, and how long the call took — all read out of the API response rather than restated from the request. The raw JSON is one click away.
That last field exists because transparency is the one parameter where asking for something and getting it are routinely different things. OpenAI's own guidance puts it plainly: a drawn checkerboard is not transparency. So the alpha channel is checked rather than assumed.
What is not fixed
The checkerboard-and-noise artifact that has followed GPT Image since the previous generation is still present in 2.5 — most visibly in clouds, grass, hair and behind dense text, and more visibly the more times an image is edited. It is widely reported and it is not a watermark; OpenAI's provenance signal is C2PA metadata and an invisible watermark, neither of which is a visible pattern.
Measurements of where it shows up, and at which settings it shows up least, are being written up separately.