GPT Image 2.5 is OpenAI’s family of image generation and editing models. The part most readers get wrong first is that it is not a model ID you can call.
This article settles that naming question first, then works through how the two variants divide the job, what the specifications allow, and a few official prompt examples you can adapt.
GPT Image 2.5 is not a single model ID
OpenAI does not publish a model called gpt-image-2.5. It publishes two, and their positioning is deliberately split.
gpt-image-2.5-flare: the default for most applications, positioned around everyday generation speedgpt-image-2.5-sunburst: positioned around editing precision, for work where detail and consistency both have to survive
This decides which parameter you pass at call time. It is also where a lot of comparison articles start going wrong, because they treat the family name as a model name and every example after that is wrong too.
Choosing between Sunburst and Flare
The difference is not price, it is the kind of task. The rates on the official pricing page are the same for both, so what drives the choice is how much waiting the job tolerates and how badly it would go to break something that already looked right.
| Dimension | gpt-image-2.5-flare | gpt-image-2.5-sunburst |
|---|---|---|
| Positioning | Default for most applications | Editing precision first |
| Priority | Speed and throughput | Detail and consistency |
| Typical work | Social assets, bulk drafts, exploring directions | Swapping backgrounds, local edits, keeping a reference |
| Quality tiers | Identical | Identical |
A practical way to decide is to write the acceptance criteria first. If “nothing that was already right may break” is hard, start with Sunburst. If you want a few dozen directions fast, use Flare first and return to Sunburst for the version you keep.
Specifications and quality tiers
Both variants share one parameter set, so switching between them means changing the model name, not rewriting the call.
- Six quality tiers:
low,medium,high,xhigh,max,auto - Recommended sizes are
1024x1024,1536x1024and1024x1536 - Custom sizes are supported, within limits on multiples, aspect ratio and total pixels
- Transparent backgrounds are supported, which suits asset and design work
- Generation uses the generations endpoint, editing an existing image uses edits
- For several rounds of changes, use the image tool on the Responses API
Check the official documentation for the exact values. These limits move, and copying a secondhand figure is riskier than reading the source once.
Core capabilities
Text-only generation covers illustrations, covers and poster backgrounds built from nothing. Reference-based editing is the focus here: you upload a product shot or a portrait and describe only the part that should change, leaving the rest to the model to hold. Inpainting repairs one region without disturbing the composition.
Text rendering improved but still has edges: poster headlines and short captions are more reliable than long passages or precise typography, which still deserve a human pass.
Three official examples and their prompts
The prompts in the official prompting guide share one habit: subject, action, lens, light, material and exclusions are written separately instead of compressed into a single adjective like “cinematic”.
Example one, photorealistic people. The official example describes an older sailor adjusting a net on a fishing boat, then names skin texture, focal length, depth of field and film grain, and closes by ruling out glamorisation and heavy retouching. Photorealism often turns on those exclusions.
Example two, exact text in an advertisement. The brand name and tagline are quoted verbatim, and the prompt asks for the tagline exactly once, integrated into the layout, with no extra text, watermarks or unrelated logos. Commercial work still needs a proofread before it ships.
Example three, an educational infographic. The official example names the audience first, then lists the molecules that must be labelled, then asks for a white background, simple icons and readable text while banning tiny type and decoration.
The full prompts are in the official prompting guide. Adapt them to your own task and test on a small batch rather than treating length as quality.
Typical use cases
Commerce is the most direct fit: white-background shots, scene images and background swaps that used to be outsourced can now be produced in batches. Social assets, article headers and campaign posters follow closely.
Design assets are the other cluster, where transparent backgrounds move cut-out work earlier into generation. Decide the review step before you start: for advertising, portraits, trademarks and news material, generation is only the first step and a human check before publication is not optional.
How to get started
Ordinary users can open ChatGPT and start, without buying a separate subscription or applying for an API key. The thing to work out first is available capacity, not whether you have access.
Developers go through the API: learn the endpoints and parameters, get one or two simple tasks running end to end, and record the usage figures the response returns.
If the real question is whether you have to pay, read free access and limits. To sketch first and generate second inside ChatGPT, see how Sketch works. Chinese typography is a separate common difficulty, covered in Chinese text rendering.
Frequently asked questions
Are GPT Image 2.5 and ChatGPT Images 2.5 the same thing? Yes, named for two contexts: inside ChatGPT it is Images 2.5, and in the API it is the two model IDs above.
Do the two variants cost the same? The rates on the official pricing page are identical, so price does not decide the choice.
Are Sketch, templates and comment annotations API capabilities? No. All three are ChatGPT product features with no equivalent API parameter.
Can I reuse code written for the previous generation? The interface shape is close, but the model ID has to change to one of these two. Run a regression test before migrating.
Does it generate Chinese text? It does, but the error rate inside the image is higher than for English. The working method is in Chinese text rendering.
Whether GPT Image 2.5 suits you depends less on how much better it is than the previous generation than on whether your task is about speed or precision. Sort the tasks by that question first, then pick the model.