GPT Image 2.5 Flare vs Sunburst: 6 Same-Prompt Tests

OpenAI shipped GPT Image 2.5 on September 8 as two API models instead of one: gpt-image-2.5-flare and gpt-image-2.5-sunburst. The developer forum post sums up the split in one line: Flare brings the quality, editing and speed improvements, and Sunburst "adds precision for detailed creative work, with longer generation times." So we know which one is slower. What the launch posts can't tell you is whether the extra wait shows up in the picture.

Both models run in the generator on this site, so I tested it the plain way: six prompts, sent word for word to both models with the same settings, each pair started at the same moment. I timed every run, then zoomed into the places image models usually slip: small type, labels, reference details, and objects that have to look like they work.

In four of the six tests I had to zoom in to find any quality difference, and in two of those there wasn't one. Where they did differ, Sunburst was better at physical detail, and Flare was better at doing exactly what the prompt asked. Sunburst was also slower on every run, by 20 to 35 seconds at 2K.

What OpenAI says GPT Image 2.5 Flare and Sunburst are for

OpenAI positions Flare as the default for most applications: higher quality than GPT Image 2 at roughly half the latency, aimed at creator content, product experiences, visual search and high-volume generation. Sunburst is for premium work that needs tighter control across edits, such as production-ready campaign creative and polished product imagery, and it takes longer per image.

For the 2.5 generation as a whole, OpenAI lists four improvements over Images 2.0: more natural lighting and richer textures, better preservation of subjects from reference photos, more reliable edit-following across multiple turns, and latency down by up to 50%. ChatGPT also gained Sketch, templates and prompt sharing, but those stay inside ChatGPT.

The tests below cover what a single image can show: lighting, texture, text, reference fidelity and a precise edit. Long multi-turn edit chains weren't part of this round.

How I ran the GPT Image 2.5 tests

• Same prompt, same settings. Every test used the same aspect ratio, 2K resolution and high quality on both models. One extra speed check at the default medium/1K tier is in the timing table.
• Same moment. Each pair went out together through the backend that runs the generator here, so both models faced the same queue.
• Wall-clock timing. Seconds from submitting the job to the finished file, checked every three seconds, so read each number as ±3 s.
• One run per model per test. None of the six 2K times looked like an outlier, so none was rerun. Another day will give other numbers.

The prompts are exactly what I sent. The images are raw outputs; the only editing is the cropping in the close-up rows.

1. GPT Image 2.5 on dense text: a night market program

A printed event program poster for a neighborhood night market. Flat graphic design on cream paper, printed in deep green and tomato red.

Title at the top: "LANTERN STREET NIGHT MARKET"
Subtitle below it: "Saturday, October 3 · 5 PM to 11 PM"

A schedule table in the middle with six rows, time on the left, event on the right:
5:00 PM | Stalls open
6:00 PM | Dumpling folding workshop
7:00 PM | Brass band on the east stage
8:00 PM | Paper lantern making for kids
9:00 PM | Tasting walk, meet at the fountain
10:30 PM | Last orders

Lower left: a simple street map with numbered dots and a legend reading "1 Food", "2 Crafts", "3 Music", "4 Restrooms".
Footer in small print: "Free entry · Cash and cards welcome · Dogs on a leash until 8 PM"

Clean grid, consistent type sizes, every word spelled exactly as written, no logos.

Settings: 2:3, 2K, high quality.

That's close to 70 words of copy. Neither model misspelled anything in the title, the six schedule rows, the legend or the footer, which is a good result for a poster this dense.

The differences are in what each model did around the copy:

• Sunburst made the nicer-looking poster and the less accurate one. It swapped the plain header for an illustration of tents, string lights and a fountain, and the date line disappeared: "Saturday, October 3 · 5 PM to 11 PM" isn't on the poster at all. It also wrote street names I never gave it onto the map ("Maple St", "Riverside Ave"), and at full size there's a grey smear across the "en" in "Stalls open".
• Flare stayed literal. Every line is there, including the date. The map icons even match the legend: food stalls at 1, a small stage with music notes at 3, a restroom sign at 4. Its only flaw is that the longest row sits tight against the right edge of the table.

For anything with copy that has to be exact, Flare won this one clearly. An event poster without its date is a reprint.

2. GPT Image 2.5 product shot: Sunburst's home turf?

Campaign hero photograph of an amber glass bottle of cold brew coffee concentrate with a brushed aluminum screw cap.
A plain kraft paper label with small black lettering that reads "COLD BREW No. 7".
The bottle stands on a wet dark slate slab, fine condensation on the glass, three ice cubes and a few coffee beans beside it.
Soft key light from the upper left, a thin rim light along the right edge of the bottle, controlled reflections, deep charcoal background.
Bottle on the right third of the frame, empty space on the left for headline copy.
Photorealistic commercial product photography, no other text, no logos.

Settings: 16:9, 2K, high quality.

Polished product imagery is the job OpenAI names for Sunburst, and at normal size I couldn't pick a winner. Same bottle, same slab, three ice cubes each, the bottle on the right third and the left side kept empty for copy.

The close-up is where Sunburst's extra time shows. Its condensation is sharper and more varied, light glows through the amber liquid, and the rim light along the right edge is clearly there. Flare's glass is darker and flatter next to it, though I'd still use it.

Flare got the label exactly right on one line: "COLD BREW No. 7". Sunburst split it over two lines and dropped the space, so it reads "No.7". Minor, but labels are what a client checks first.

If the shot will run large, as a banner or a landing page hero, Sunburst is worth the extra 24 seconds. In a grid of small product thumbnails, nobody will see the difference.

3. Light and texture in GPT Image 2.5: a woodworking bench

Late afternoon in a small woodworking shop. Two hands push an old cast-iron hand plane along a walnut board, and a long curl of wood shaving rises from the blade.
Low golden sunlight comes through a dusty window on the left: a visible light beam full of floating sawdust, warm highlights on the walnut grain, soft deep shadows under the bench.
Show these textures clearly: open wood grain, worn cotton sleeves, scratched metal, rough plaster wall.
Medium close-up at bench height, shallow depth of field, photorealistic, no faces, no text, no logos.

Settings: 3:2, 2K, high quality.

At first glance these are the same photo: warm light from the left, sawdust in the beam, walnut grain, frayed sleeves. Flare's light beam is the more dramatic of the two, with more dust hanging in it.

Then look at the tool. Sunburst drew a hand plane that could actually work, with a front knob, a lever cap with a brass adjuster, and a shaving curling up out of the mouth ahead of the blade. Flare's plane is harder to read. The shaving and the lever run into each other, and there's no proper rear handle, so the back hand grips bare metal.

Nobody scrolling past will notice. A woodworker will, and so will anyone selling tools. This is the pattern I kept seeing: Sunburst gets physical objects more right, and you usually need a close-up to see it.

4. GPT Image 2.5 editing: three changes, nothing else

Reference image: the desk photo from the GPT Image 2.5 page.

Edit this photo. Change only three things:
1. Make the white mug matte black.
2. Replace the succulent with a small round cactus in a terracotta pot.
3. Remove the reading glasses completely.
Keep everything else exactly as it is: the desk, the grey notebook, the wall texture, the lighting, the shadows and the camera angle.

Settings: 4:3, 2K, high quality, one reference image.

A draw. Both made all three changes and left the rest alone. The mug kept its shape and square handle, now matte black. The glasses are gone, with the wood grain underneath filled in cleanly. The notebook, the wall, the shadows and the camera angle match the original in both.

The one difference is the pot. Flare used a weathered clay pot with white mineral marks. Sunburst kept the original pot's tapered shape and speckled glaze and recolored it terracotta, which stays closer to the photo.

OpenAI pitches Sunburst for tighter control across edits. One edit didn't show a gap, though a long chain of edits might, and I haven't tested that yet. For a single edit like this, Flare got the same result 20 seconds sooner.

5. GPT Image 2.5 with three reference images

Reference images: a knitted fox, a skincare bottle, and the desk from test 4, all from the showcase images on the GPT Image 2.5 page.

Use the three reference images.
Image 1: a hand-knitted plush fox. Image 2: a frosted glass bottle with a sage-green cap and a label that reads "FIELD & FERN". Image 3: an oak desk against a pale plaster wall.
Create one new photo: the same fox sits on the same oak desk, leaning against the same bottle.
Keep the fox's one folded-over ear, black button eyes and yarn texture, the bottle's cap color and label text, and the desk's wood grain and wall.
Soft morning window light from the left, eye-level camera, photorealistic, no other text, no logos.

Settings: 1:1, 2K, high quality, three reference images.

Reference fidelity is one of the upgrades OpenAI claims for 2.5, and both models delivered it. The fox kept the folded ear on the correct side, the button eyes, the cream chest and the dark paws. The bottle kept its sage cap and a correctly spelled "FIELD & FERN" label. The desk kept its oak grain and plaster wall.

They read "the same oak desk" differently. Flare took the desk and the wall and left the props behind, which gives a clean two-subject frame. Sunburst brought the succulent, the mug and the notebook over from the desk photo, and rested the fox's head right on the bottle cap. That's more literal, and also more clutter if you wanted a product shot.

Neither is wrong. If you want only the subjects you named, say it in the prompt or use Flare. At 82 seconds, this Sunburst run was the slowest of the whole test.

6. Chinese text in GPT Image 2.5: a tea shop menu

A minimalist menu board for a small tea shop, printed on warm off-white card in ink black and celadon green.
Title in Simplified Chinese at the top, exactly: "山间茶铺"
A small line under the title: "每日 10:00 – 21:00"
Five menu rows, tea name on the left, price on the right:
"白毫银针" ¥38
"凤凰单丛" ¥42
"桂花乌龙" ¥32
"陈年普洱" ¥45
"冷泡绿茶" ¥28
Footer line: "续水免费 · 茶叶可外带"
Elegant Song-style Chinese serif, generous spacing, every character exactly as written, no other text, no logos.

Settings: 2:3, 2K, high quality.

Both got every character right: the shop name, all five tea names, the prices, the opening hours and the footer. Dense characters like 毫 and 凰 hold their strokes at full size in both.

Only the layout differs. Sunburst added thin rules between the rows, small leaf ornaments and a faint ink-wash mountain; Flare kept the menu open with leaves in two corners. I slightly prefer Sunburst's menu, but on Chinese text itself this was a tie.

GPT Image 2.5 Flare vs Sunburst speed, test by test

On every 2K high-quality run, Sunburst took between 20 and 35 seconds longer, about 1.6 times Flare's time on average. At the default medium/1K tier the gap came down to 10 seconds. Your numbers will move with settings and load, but in my runs the order never changed.

Price doesn't separate them. On this site both models use the same credit schedule at every quality and resolution: 8 credits an image at the default tier, less at low quality, more at higher quality or resolution. So the choice comes down to time against the details above.

So which GPT Image 2.5 model should you use?

My default is Flare, with Sunburst for the final render of a hero image. That's close to OpenAI's own positioning, though not quite for the reason I expected. Sunburst didn't follow prompts more closely. It drew glass, light and objects more convincingly.

GPT Image 2.5 API model IDs and settings

If you're calling the API yourself, these are the model names:

Use the dated snapshot when the model behind your prompts needs to stay fixed.

The parameters worth knowing, from the image type definitions in the openai-python SDK:

• quality: low, medium, high, xhigh, max or auto. xhigh and max exist only on the 2.5 models. The generator here stops at high, so I haven't tested either.
• size: auto, 1024x1024, 1536x1024, 1024x1536, or any WIDTHxHEIGHT where both sides divide by 16 and the ratio sits between 1:3 and 3:1, up to 3840x2160. The SDK marks sizes above 2560x1440 as experimental.
• background: opaque, transparent or auto. Transparent needs png or webp output.
• output_format: png, jpeg or webp, with output_compression from 0 to 100.
• moderation: low or auto. n: 1 to 10 images. With stream, partial_images goes from 0 to 3.
• Edits accept up to 16 input images (png, webp or jpg, each under 50 MB), an optional PNG mask under 4 MB, and input_fidelity set to high or low.

A minimal Python call:

import base64
from openai import OpenAI

client = OpenAI()

result = client.images.generate(
    model="gpt-image-2.5-flare-2026-09-08",
    prompt="An amber glass bottle of cold brew on wet slate, thin rim light",
    size="1536x864",
    quality="high",
)

with open("cold-brew.png", "wb") as f:
    f.write(base64.b64decode(result.data[0].b64_json))

Change the model string to gpt-image-2.5-sunburst-2026-09-08 to send the same request to Sunburst. Launch details are in OpenAI's announcement and the developer forum post.

How to run the Flare vs Sunburst test yourself (3 steps)

1. Open GPT Image 2.5 or the text-to-image generator. Flare and Sunburst are both in the model menu, you don't need a ChatGPT account, and new accounts start with free credits.
2. Run one prompt on both models with identical settings. Copy any prompt above. At the default medium/1K tier each image costs 8 credits on either model. For tests 4 and 5, use the image-to-image generator, which takes up to 16 reference images.
3. Zoom in before you decide. Check the smallest text, every label, and whatever has to make physical sense. If you can't see a difference at the size the image will actually be used, stay on Flare.

The Bottom Line

Leave Flare selected. It kept every line of copy, matched Sunburst on the edit, the references and the Chinese menu, and finished 2K images about 26 seconds sooner on average. Sunburst's lead is real but narrow: glass, light and objects that survive a close-up. Save it for the final render of an image people will look at closely, and proofread anything it adds on its own.

Related reading

• 10 GPT Image 2 Poster Prompts I Use for Client-Ready Designs
• 5 GPT Image 2 Cinematic Lighting Prompts
• GPT Image 2 Style Library: 12 Copy-Paste Art Style Prompts

gpt-image2.art is an independent site and is not affiliated with OpenAI. Model names are used only to identify the models tested.