GPT Image 2.5

You can generate an image from a description and edit a photo. ChatGPT offers limited free access; you cannot download the model itself to your computer.

A fit if you are exploring ideas for a page, illustration, or ad image and are ready to choose and refine a result.

Not a fit if, after replacing one object, every other pixel must remain unchanged.

For a first look, open ChatGPT. To choose between the two options, go to the Flare and Sunburst comparison.

The capabilities are clearest on a task with a source image. Below are published OpenAI examples and one OneWave analysis. We compared the results; this profile contains no generations of our own with this model.

App mockup. With a single farmers market prompt, both models create a screen with products and photos. Flare has one large product block, while Sunburst has several cards. This can be passed to Astra as a visual reference for implementation.

From drawing to landscape. A river, mountains, and a tree move from a sketch into a realistic valley. The horizontal source image becomes vertical. The composition's meaning is preserved better than its exact proportions.

Clothing from references. A jacket, top, and boots are attached to a photo of a woman. Her face, pose, and the museum remain recognizable. Notice the jeans: Flare lightens them, while Sunburst keeps a shade closer to the original.

Furniture in a room. White chairs are replaced with wooden ones; the kitchen remains recognizable by its windows, light fixture, and refrigerator. But the framing and aspect ratio change. That is enough to align on the idea; precise photo compositing requires verification.

Long brief. OneWave tested a scene with books, a mug, and specified lighting. The author gave both models credit for all eight requirements. This is a useful separate example of prompt following; the author did not publish a complete log of API calls.

The official user entry point is ChatGPT. Choosing Flare or Sunburst by name requires the OpenAI API. The ChatGPT profile covers the app's plans and capabilities.

API names: gpt-image-2.5-flare and gpt-image-2.5-sunburst. For a reproducible version, you can pin a snapshot ending in -2026-09-08. Organization verification may be required.

Start with one clear change: "replace the white chairs with wooden ones, keeping the rest of the room unchanged." State what may change and what must remain the same. Then compare the source image and result before requesting the next edit.

Russia is not on OpenAI's official list of supported countries. Access to ChatGPT and access to the API should be checked separately; a chat subscription does not pay for programmatic requests.

In the first Arena tables, both models ranked above GPT Image 2. These are still preliminary results: the new models have noticeably fewer votes, and quality settings are not specified.

Elo reflects participant preferences in pairwise comparisons. In text-to-image generation, the Flare and Sunburst intervals overlap, so the top positions do not yet provide a confident answer about which is better for your prompt. A series of your own typical tasks is more useful than a single overall rank.

Flare and Sunburst have the same token rates. This is not a fixed image price: usage depends on input images, output size, and quality.

Quality levels: low, medium, high, xhigh, max; auto lets the model choose the setting. For your budget, check usage on real prompts.

OpenAI's claim that Flare is faster refers to a comparison with GPT Image 2. There is no universal time for every prompt. Costs work differently in ChatGPT: plan terms and app limits apply.

When editing, you can get the desired change while quietly losing an important detail. Translating a coffee machine diagram shows why even a good-looking result still needs to be checked against the original.

The request concerned language, but Flare changed the part number. Such an error is more dangerous than an obvious typo: the diagram's overall look makes it seem that everything stayed in place.

For client source files, check storage terms. In the API, data is not used for training by default, but operational logs can generally be retained for up to 30 days. These terms do not apply to personal ChatGPT.

For a first illustration option, I would start with Flare. If preserving a character or details through successive edits matters more, I would test the same prompt in Sunburst.

Test both on the same source image, prompt, and quality setting. Measure the time to an accepted image, including corrections. Savings from a fast first attempt disappear if the result needs reworking several times.

To implement the drawn screen, you need a coding agent: for example, Codex with Astra. GPT Image handles the image, while the app connects it with the rest of the work.