ChatGPT runs GPT Image 2.5 but won't tell you which of its two models answered, and won't let you pick. Here it's a toggle, and every image comes back with the model ID printed under it. Edits re-render from your original upload, never from the last output, so a fourth pass is as clean as the first. First image free, no account, no card.
OpenAI shipped GPT Image 2.5 on 8 September as two separate models — Flare, which matches GPT Image 2 at about half the latency, and Sunburst, which is the one that actually beats it. Inside ChatGPT you can't choose between them, and you can't find out which one drew your picture.
So the first thing gptimage25.top does is print the answer: under every result, the model ID the API returned, the quality tier, the pixel size and the elapsed time, with the raw JSON one click away.
The second thing is editing. Chat editors re-edit their own last output, which is why quality falls apart by the third pass. Here you tick what must not change — face, camera angle, text, lighting, product shape — and it re-renders from the file you uploaded.
First image each day is free, no account, no card. Failed and moderation-blocked runs never cost credits.
Two things it does not do, on the pages rather than in the small print: no transparent PNG yet, and no cross-session identity lock. Where Nano Banana Pro is better — large soft textures, photorealistic portraits — the comparison page says so first.
Not affiliated with OpenAI. Built by one person.