GPT Image 2.5: Two Models, and Edits That Finally Stick
OpenAI shipped two image models at once. Flare is fast, Sunburst is careful, and both make an edit mean an edit. Here's what changed since GPT Image 2 and which one to run on each::labs.
You ask for one small change. Swap the mug for a glass, warm up the light, fix the label. The model hands back a new picture. The glass is there, sure, but the hand holding it has a different ring, the window moved, and the label now reads like a ransom note. You didn't ask for a new image. You asked for an edit.
That gap between "change this" and "change everything" is the quiet tax on every image workflow. It's the problem GPT Image 2.5 is pointed at. OpenAI shipped two models this time, Flare and Sunburst, and both are live on each::labs with Text to Image and Edit endpoints. The split says more about where image generation is heading than any sample image could.
What GPT Image 2.5 Actually Is
GPT Image 2.5 is OpenAI's newest image generation and editing family, released on September 8, 2026, according to OpenAI's developer changelog. It follows GPT Image 2, which landed in April, and it takes text and images in and gives images out. No video, no audio. It's a still-image model that wants to be very good at stills.
The headline isn't a single capability. It's the structure. Instead of one model with a quality dial, you get two models built for different jobs, and OpenAI is unusually direct about which is which. Its own announcement frames the release around "higher image fidelity, stronger style adherence, and more control over edits," and that last phrase is the one to underline. Generation got better. Editing is what got serious.
On each::labs that becomes four endpoints: Flare Text to Image, Sunburst Text to Image, and an Edit model for each. Same inputs on both sides, so switching between them is a model change and nothing else.
Flare and Sunburst Are OpenAI's Names, Not Ours
Worth clearing up first, because the names sound like a marketing team got bored: Flare and Sunburst are OpenAI's official model names. The API IDs are gpt-image-2.5-flare and gpt-image-2.5-sunburst, and that's what each::labs runs under the hood.
OpenAI's image prompting guide spells out the difference in two sentences. Flare is "the small model, optimized for speed, with image quality comparable to GPT Image 2." Sunburst is "the base model, optimized for quality, with higher image quality than GPT Image 2." Its model pages call Sunburst the most capable model for generation and editing, and Flare the fastest for high-quality everyday images.
So it isn't good and better. It's fast and careful. Flare gives you the quality you trusted from GPT Image 2, quicker. OpenAI's launch post goes a step further and claims Flare delivers higher quality at 50% lower latency than GPT Image 2; the docs are more conservative and say "comparable." Treat the docs as the floor and test for the rest. Sunburst trades speed for a higher ceiling, and the each::labs page for it says exactly that: longer generation times in return for higher precision.
What Changed Since GPT Image 2
If you've been running GPT Image 2 on each::labs, a few differences show up the moment you open the request schema.
Transparent backgrounds are in. GPT Image 2 didn't support them, and its each::labs schema says so outright. With GPT Image 2.5, both models accept background set to transparent, as long as you output PNG or WebP. For anyone making logos, stickers, product cutouts or UI assets, that one change removes an entire cleanup step.
Quality goes past high. GPT Image 2 topped out at high. Both 2.5 models add xhigh and max settings on top of low, medium and high, and OpenAI confirms the earlier models stop at high. Expect them to slow things down. They're the settings you reach for on a final asset, not on twenty drafts.
Editing gets a real mask. The 2.5 Edit endpoints on each::labs take an optional mask image, a PNG with an alpha channel at the same size as your first input, where fully transparent pixels mark the area to change. The GPT Image 2 Edit endpoint here has no mask field at all.
Bigger canvases feel less experimental. Sizes still run from 1024 square up to 4K (3840 by 2160 or the vertical equivalent), with custom dimensions in multiples of 16 and aspect ratios between 1:3 and 3:1. The difference is in the fine print: GPT Image 2 labeled its 2K presets experimental, while for 2.5 only resolutions above 2560 by 1440 carry that warning.
The look moved too. OpenAI promises more natural lighting, richer textures and better handling of complex layouts. Verify that on your own prompts.
The GPT Image 2.5 Editor: Change One Thing, Keep Everything Else
This is the section that matters if you searched for a "GPT image editor" and landed here. Editing is where GPT Image 2.5 earns its version number.
Both models, per OpenAI, "offer improvements in precise editing and subject preservation." The each::labs descriptions get specific. Flare Edit targets only the requested changes while preserving the subject, composition and background, with consistent identity across styles and multiple rounds of edits. Sunburst Edit is built to apply changes exactly where requested and keep the subject, structure and composition consistent through repeated revisions.
Notice the emphasis isn't on the first edit. It's on the fifth. One convincing change is easy. Keeping a face, a product's geometry and a background intact through a chain of revisions is the hard part, and it's exactly what OpenAI points at when it calls Sunburst the model for consistency across edits.
The input side is generous. Each Edit request takes up to 16 source images, so you can pass a person plus clothing references, or a scene plus the object you want placed into it, and tell the model which image plays which role. Add the mask when you need a local change and want the rest of the frame left alone.
One caution from OpenAI's own docs: masking with GPT Image is prompt-based. The model uses the mask as guidance and may not follow its exact shape with complete precision. Think of it as a strong hint, not a stencil.
Pick Flare First, Earn Your Way to Sunburst
The tempting move is to default to the "most capable" model and never think about it again. Don't. OpenAI's own migration advice runs the other direction, and it's the right call.
If GPT Image 2 already meets your quality bar, start with Flare and look for the speed gain. If GPT Image 2 falls short on a complex job, start with Sunburst, confirm it actually clears the bar, then run the same prompts and inputs through Flare. Switch only if the quality holds. Keep Sunburst only where its advantage is visible in the output, not in the spec sheet.
In practice the split is predictable. Flare suits high-volume work: social content, product listing variants, rapid concept rounds, anything where you'll look at forty images and keep three. Sunburst suits the jobs where a reviewer will zoom in: hero images, campaign key visuals, packaging with real copy, and multi-step edits where a face or a product has to survive every round.
A natural workflow uses both. Explore on Flare at medium quality, pick the direction, then finish on Sunburst at xhigh or max. Because the two share the same request shape on each::labs, that hand-off is one string in your code.
A Few Prompts to Feel the Difference
This isn't a prompting guide (that one's coming next in this series), but a handful of test prompts will show you the gap between the two models faster than any description. Run each on both and compare.
A transparent asset, the feature GPT Image 2 couldn't do. Set background to transparent and output to PNG:
A flat vector badge for a neighborhood bakery: a round loaf with a wheat stalk crossing it, warm ochre and cream palette, thick clean outlines. Single centered mark, generous padding, fully transparent background, no shadow, no scenery.
A layout with exact text, the classic stress test:
A minimal concert poster, portrait. Deep navy background, a single brass trumpet drawn in fine gold line art. Title text exactly "Late Set", set once in a tall condensed serif at the top. No other text, no watermark.
A local edit on Flare Edit or Sunburst Edit, with a mask over the mug only:
Replace the coffee mug with a clear glass of iced tea with lemon. Keep the hand, the table, the window light and everything else exactly as it is. Match the shadows and color temperature of the original photo.
A multi-reference edit, passing a portrait as image 1 and a jacket as image 2:
Dress the person from image 1 in the jacket from image 2. Do not change their face, hair, pose, background or framing. Fit the fabric naturally to the body and match the lighting of image 1.
Then do the test that actually separates them: take the output of that last edit and ask for a second change, then a third. That's where you'll learn whether you need Sunburst.
Who Should Move to GPT Image 2.5 Now
If you're on GPT Image 1, 1.5 or 1 Mini anywhere in your stack, the clock is already running. OpenAI's deprecations page lists gpt-image-1 for shutdown on October 23, 2026, and gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest for removal from its API on December 1, 2026. The recommended replacement for every one of them is Sunburst or Flare. That's not a nudge. It's a migration deadline.
If you're on GPT Image 2 and happy, there's no emergency, but there's an easy experiment waiting: run your production prompts through Flare and see whether you get the same images sooner. If you've been hand-removing backgrounds or fighting edits that drift, move today. Those are the two gaps this release goes after most directly.
New to the family? Compare it with the other text-to-image models on each::labs. For raw speed at volume, our look at Nano Banana 2 Lite is a useful counterpoint. For context on where this family started, here's the day GPT Image first arrived on each::labs.
Wrapping Up
GPT Image 2.5 isn't a single leap. It's a fork in the road, and OpenAI has labeled both paths clearly. Flare is the everyday engine: GPT Image 2 quality, faster, now with transparency and higher quality settings. Sunburst is the careful one, the model you hand a hero image or a fifth-round revision and trust not to wander.
The real upgrade is that an edit can finally mean an edit. That sounds modest until you count the hours spent fixing what a model changed without being asked.
Both models are live on each::labs right now. Start on Flare, keep Sunburst for the jobs that earn it, and watch for the Prompting Guide next in this series.
Frequently Asked Questions
What's the difference between GPT Image 2.5 Flare and Sunburst?
Speed versus ceiling. Flare is OpenAI's small model, tuned for speed with quality comparable to GPT Image 2. Sunburst is the base model, tuned for quality above GPT Image 2. Both generate and edit, and both take the same inputs on each::labs, so start with Flare and move up when edits must hold across many rounds.
Is GPT Image 2.5 good for image editing?
Editing is the main reason it exists. OpenAI says both models improve precise editing and subject preservation, and the Edit endpoints on each::labs accept up to 16 reference images plus an optional inpainting mask. Masks act as guidance rather than an exact stencil, so check edges on anything you'll composite.
Does GPT Image 2.5 support transparent backgrounds?
Yes, and it's one of the clearest upgrades over GPT Image 2, which didn't. Set background to transparent and choose PNG or WebP output. It works on both Flare and Sunburst.
When does OpenAI shut down GPT Image 1?
Sooner than many people expect. OpenAI's deprecations page lists gpt-image-1 for shutdown on October 23, 2026, with gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest following on December 1, 2026. OpenAI recommends GPT Image 2.5 Sunburst or Flare as the replacement.