# GPT-Image-2.5 Sunburst edges Flare and GPT-Image-2 on our ten prompts

Prompt and Render desk. A ZeroClick publication (https://promptandrender.com/about/). Published 2026-10-20.
Format: Model test. Topic: Tools and platforms (https://promptandrender.com/topics/tools-and-platforms/).
Canonical: https://promptandrender.com/articles/gpt-image-2-5-flare-sunburst-vs-gpt-image-2/

OpenAI released two new image models for its API on 8 September 2026, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. According to its [launch announcement](https://openai.com/index/introducing-chatgpt-images-2-5/), the new generation brings more natural lighting and richer textures, more accurate content in images that draw on real-world information, more complex layouts including transparent backgrounds, and better fidelity to visual styles. OpenAI pitches Flare as the default for most applications, with higher quality than GPT-Image-2 at half the latency. It positions Sunburst as the more precise option for detailed creative work, with longer generation times.

We wanted to see those claims for ourselves, so we ran both new models and GPT-Image-2 through a set of ten reference prompts. Each prompt is designed to test specific, noticeable details: steam lit from behind, a headline spelt exactly, laces against an empty ground, three cats who must each look somewhere else, and a weather screen that has to say °C.

We sent each prompt once to each model through OpenAI's Image API, at 1024x1024, medium quality and PNG, with costs estimated at OpenAI's list rates; every image below was generated by the AI model named under it.

## Prompt 1: backlit steam in low sunset light

Lighting heads OpenAI's list of improvements, and steam is a demanding test of it: there is almost nothing there to light, so the glow has to come from the window behind it. The low sun has a second job, raking across an old oak table to pull out the grain.

```text
A ceramic teapot on a weathered oak table by a west-facing window at sunset, low orange light raking across the table, steam lit from behind, shallow depth of field, shot on 50mm film.
```

![Blue-and-white patterned teapot on a worn wooden table, a column of steam glowing orange against a window with the sun setting behind it.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/3364fbf035c242fe/display.webp "GPT-Image-2.5 Flare")
![Floral teapot on a rustic table beside a cup, flowers and a plaid cloth, a faint wisp of steam and the setting sun in the window behind.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/6b5b66dd6b2a095c/display.webp "GPT-Image-2.5 Sunburst")
![Plain speckled stoneware teapot on rough planks with a long shadow, orange-lit steam rising against a dark room, the sun low in the window at left.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/74df49a1c7c396c6/display.webp "GPT-Image-2")

*Prompt 1: look at the steam against the window, and at how the light crosses the grain of the table. 1024×1024, medium, 2026-10-19.*

Flare makes the steam the subject: a glowing column rises against the window while orange light pools in the cracks of the oak. GPT-Image-2 gives the most filmic frame, a plain pot casting a long raking shadow, with its steam also lit from behind. Sunburst dresses the table best but has the faintest steam of the three. Flare and GPT-Image-2 are too close to call, so this prompt doesn't prove OpenAI's lighting claim.

## Prompt 2: exact text on a poster

Lettering has long been where image models give themselves away, with headlines that come out as near-words and letters that melt halfway through a line. OpenAI's own [image generation guide](https://developers.openai.com/api/docs/guides/image-generation) still lists text rendering as a limitation, so this prompt calls for a poster with an exact headline, a date line with a middle dot, one illustration and nothing more.

```text
A minimalist concert poster with the exact headline "NORTH SHORE JAZZ NIGHT", the line "Friday 7 PM · Harbor Hall", and one brass trumpet illustration, navy ink on cream.
```

![Cream poster with a navy serif headline over three lines, the middle word set large, an engraved trumpet below and the date and venue line at the foot.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/9dfcb6f3f9584651/display.webp "GPT-Image-2.5 Flare")
![Cream poster with a two-line navy headline, an engraved trumpet, and the date and venue line between two thin rules.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/67a2c69c68d7ad5b/display.webp "GPT-Image-2.5 Sunburst")
![Cream poster inside a navy border with a condensed two-line headline, an ornamental rule, an engraved trumpet and the date and venue line between rules.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/f97a7807ef7ad01f/display.webp "GPT-Image-2")

*Prompt 2: check every letter, the middle dot in the date line, and how much ornament each poster adds. 1024×1024, medium, 2026-10-19.*

All three models set every word right, middle dot included. With the text a tie, the posters come down to restraint. GPT-Image-2 adds a border and an ornamental rule, which makes it the busiest poster for a brief that asked for minimalist. Flare's oversized JAZZ and Sunburst's clean two-line stack both read as posters you could send to print.

## Prompt 3: a five-step infographic

OpenAI pitches the new models at "presentation visuals that fit a defined structure", and an infographic is exactly that kind of work. This one asks for a title, five steps in order and five single words, which gives a model plenty of chances to drop a word or shuffle a step.

```text
An infographic titled "How a Coffee Bean Becomes Espresso": five numbered steps left to right, each an icon and one word: Harvest, Dry, Roast, Grind, Brew.
```

![Infographic with a two-tone serif title and five tinted panels numbered one to five, each an illustrated coffee step with its word, joined by arrows.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/623c07f8827ca7fb/display.webp "GPT-Image-2.5 Flare")
![Infographic with a bold title and five round icon badges numbered one to five, joined by arrows with a word under each, and empty space below.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/e96097be74bb67df/display.webp "GPT-Image-2.5 Sunburst")
![Infographic with a serif title, five numbered illustrations from coffee cherries to an espresso shot, and the step words joined by arrows along the bottom.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/9beb22d96e19adf4/display.webp "GPT-Image-2")

*Prompt 3: look at the title, the five words, and whether each step is an icon or a picture. 1024×1024, medium, 2026-10-19.*

All three spelled the title and all five words correctly and numbered the steps from left to right. Sunburst is the only one that drew icons in the infographic sense: five round badges you could shrink onto a slide. It does leave the bottom third of the square empty. Flare and GPT-Image-2 draw small illustrations instead, and Flare's tinted panels make the richest page. All three hit; take Sunburst for a deck and Flare for a wall.

## Prompt 4: a product cut-out on a transparent ground

A product shot is most useful as a cut-out that drops onto any page, and transparent backgrounds are among the layouts OpenAI says the new models now handle. A worn glove makes the edges work hard: the laces, the frayed leather and the gaps in the webbing all have to come through as empty ground.

```text
A worn leather baseball glove with a baseball in the pocket, studio product shot, on a transparent background.
```

![Dark, well-worn brown leather glove with a scuffed baseball in the pocket, cut out on a transparent background.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/b1972f21fa3cf2c0/display.webp "GPT-Image-2.5 Flare")
![Lighter tan leather glove with a baseball in the pocket and visible laces, cut out on a transparent background.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/114b7a5ab30c838c/display.webp "GPT-Image-2.5 Sunburst")

*Prompt 4, sent with the API's transparent-background setting: look at the laces and the glove's edge against the empty ground. 1024×1024, medium, 2026-10-19.*

Flare and Sunburst both returned a cut-out glove, and both responses reported a transparent background. They drew nearly the same glove, down to the pose and the angle of the ball, like two finishes on one design. Flare's leather is darker and more scuffed, closer to the worn glove the prompt asked for, while Sunburst's looks newer; in both, the laces read cleanly and the gaps in the webbing come through as empty ground. GPT-Image-2 is a miss: the API refused the request, returning an error that said transparent backgrounds are not supported for this model. On this prompt, that launch claim holds up. The guide documents transparent PNG or WebP output for both 2.5 models, and for anyone shipping product cut-outs that settles the choice.

## Prompt 5: an Impressionist style, with no lettering

OpenAI says the new models are "better at reflecting visual styles", so this prompt names a period style rather than any one artist's: a 19th-century French Impressionist oil, set in a snowy Rome. Paintings in that style usually carry a signature in a lower corner, so the prompt also asks for no signature or lettering, to see whether a model leaves out what the style tends to include.

```text
The ruins of the Roman Forum at dusk after a light snowfall, broken columns against a pale winter sky, painted as a 19th-century French Impressionist oil: loose, visible brushstrokes, broken color and soft winter light, with no signature or lettering.
```

![Impressionist-style painting of snow-covered Roman ruins at sunset with heavy broken brushstrokes and a small mark like a signature in the bottom-right corner.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/4a2ed27f9596cd94/display.webp "GPT-Image-2.5 Flare")
![Painting of snowy Roman ruins with columns at left under a pink and orange sunset sky, in tighter, detailed strokes, with no lettering.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/047c14ff9aa63aaa/display.webp "GPT-Image-2.5 Sunburst")
![Soft, muted painting of three columns at right and a domed church at left under a violet dusk sky, with a light dusting of snow.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/ea166536cf3e383e/display.webp "GPT-Image-2")

*Prompt 5: look at the brushwork, the depth of the snow, and the bottom-right corner. 1024×1024, medium, 2026-10-19.*

Flare's brushwork is the loosest and most broken of the three, the closest to the oil the prompt describes, and it also put what looks like a signature in the corner the prompt kept clear. Sunburst paints in tighter strokes and leaves the corner clean. GPT-Image-2 is the softest and quietest, and it is the only one whose snow reads as a light fall rather than a blanket. Sunburst and GPT-Image-2 hit. Flare is a partial, because someone would have to retouch that corner before it ships.

## Prompt 6: counting and placement

Counting is a quiet weakness of image models: ask for seven of something and six or eight come back. OpenAI's guide also warns that its models "may have difficulty placing elements precisely", so this prompt adds one larger marble that has to sit behind a particular blue one.

```text
Exactly seven blue marbles in a row on a white table, with one larger green marble behind the third blue marble from the left.
```

![Seven glossy cobalt marbles in a row on white, with a larger clear green marble sitting behind the third from the left.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/e4b0f96f4bc0bd08/display.webp "GPT-Image-2.5 Flare")
![Seven blue glass marbles in a row on white, with a larger swirled green marble behind them, just right of the third from the left.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/3b9f150948016ef9/display.webp "GPT-Image-2.5 Sunburst")
![Seven blue swirled marbles in a row on a white table, with a larger green marble set back behind the third from the left.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/aff980b6dcc6e5d8/display.webp "GPT-Image-2")

*Prompt 6: count the blue marbles, then find which one the green marble sits behind. 1024×1024, medium, 2026-10-19.*

Every model counted seven blue marbles. Flare and GPT-Image-2 set the larger green one squarely behind the third; Sunburst's sits a touch to the right of it, close enough to count. That makes it a three-way tie: three hits and no winner.

## Prompt 7: two hands at work

Image models have long struggled, often comically, with hands: extra fingers, fused knuckles, thumbs on the wrong side. Kneading presses two hands against each other and into the dough, and that contact is where those mistakes tend to show.

```text
Close-up of two hands kneading bread dough on a floured wooden counter, flour on the fingers, warm kitchen light.
```

![Two floured hands pressing into a large ball of dough on a dusted wooden counter, with a kitchen, a jar and a bowl softly blurred behind.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/67ed245f8a1637af/display.webp "GPT-Image-2.5 Flare")
![Two floured hands folding dough on a floured wooden counter, with a bowl, a plaid cloth and a plant in a warm, blurred kitchen behind.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/65a29802c6c21b74/display.webp "GPT-Image-2.5 Sunburst")
![Tight close-up of two floured hands gripping and folding dough on a wooden board, filling most of the frame in warm light.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/8d9dd459353d9082/display.webp "GPT-Image-2")

*Prompt 7: look at the fingers, the flour on them, and how close each frame really is. 1024×1024, medium, 2026-10-19.*

All three show two hands, flour on the fingers and on the wood, and warm light, and no render shows an extra or fused finger. GPT-Image-2 is the only true close-up, with hands and dough filling the square. Flare and Sunburst pull back to show the kitchen. All three hit, and GPT-Image-2 wins by a knuckle on the word close-up.

## Prompt 8: a labelled diagram from the real world

OpenAI says images that include real-world information now have more accurate content. A diagram of the planets is about as easy to check as real-world information gets: eight names in one right order.

```text
A clean labeled diagram of the eight planets in order from the Sun, each planet's name beneath it.
```

![Dark diagram with a title, the Sun at left, and the eight planets in a row from Mercury to Neptune, each named underneath.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/d80edbfe42fe7e9d/display.webp "GPT-Image-2.5 Flare")
![Starfield diagram with the Sun at left and the eight planets in order from Mercury to Neptune, small labels beneath each.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/1263b2592be902db/display.webp "GPT-Image-2.5 Sunburst")
![Starfield diagram with the Sun at left and the eight planets in order, each with a bold white name beneath it.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/c6ae77d08f4f1fcc/display.webp "GPT-Image-2")

*Prompt 8: read the order and the spelling of every label. 1024×1024, medium, 2026-10-19.*

All three put the planets in the right order with each name spelt correctly beneath it. Flare added a title nobody asked for, GPT-Image-2's bold labels are the easiest to read at a glance, and Sunburst's are the smallest. All three hit. GPT-Image-2 already got this right, so the new models had nothing to improve on here.

## Prompt 9: three subjects who must differ

With several subjects in one frame, details tend to drift between them: one cat's markings turn up on another, or the whole row looks the same way. This prompt calls for three different coats, one pair of blue eyes and three different gazes, to see whether each cat stays its own.

```text
Three cats on a windowsill: a small brown tabby, a grey and white tuxedo cat, and a light grey and brown Siberian cat with blue eyes, each looking a different way.
```

![A brown tabby, a grey and white tuxedo cat and a fluffy blue-eyed Siberian on a white sill; the tabby looks off to the left while the tuxedo and the Siberian both face the camera.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/06cccca465f2ee3e/display.webp "GPT-Image-2.5 Flare")
![A tabby gazing upward, a tuxedo cat and a long-haired blue-eyed Siberian on a sill; the tuxedo and the Siberian both look off to the right.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/6a81fb18eea3391c/display.webp "GPT-Image-2.5 Sunburst")
![A tabby in profile facing left, a tuxedo cat looking up, and a blue-eyed Siberian looking toward the camera, side by side on a sill.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/535240763fbbd8f4/display.webp "GPT-Image-2")

*Prompt 9: follow each cat's gaze. 1024×1024, medium, 2026-10-19.*

All three got the three coats and the blue eyes right, so the gaze decides it. GPT-Image-2's cats look three ways: the tabby left in profile, the tuxedo up, and the Siberian toward the camera. Flare's tabby looks off to the left, but its tuxedo and Siberian both look straight at us, and Sunburst's tuxedo and Siberian both look off to the right. GPT-Image-2 wins, and both new models are partials.

## Prompt 10: an interface that names its units

OpenAI also pitches the models at "UI concepts that preserve a given hierarchy". A weather screen is a fair stand-in: a header, five days, an icon for each, and a temperature that carries its unit.

```text
A flat-design phone screen for a weather app: header "This week", a five-day forecast with icons and temperatures in °C.
```

![Phone mockup with a blue sky header reading This week, five day cards from Monday to Friday with weather icons and temperatures shown only as degrees.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/68c2a8754f59d792/display.webp "GPT-Image-2.5 Flare")
![Phone mockup with a This week header over hills, five rows from Monday to Friday with icons and Celsius highs and lows, and a tab bar below.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/d42fa5a995b8f317/display.webp "GPT-Image-2.5 Sunburst")
![Full-bleed weather screen headed This week, five dated rows with icons, condition words and Celsius highs and lows.](https://cdn.zeroclick.io/hub/prompt-and-render/uploads/e61a65b2f24fe414/display.webp "GPT-Image-2")

*Prompt 10: read the temperatures and look for the C. 1024×1024, medium, 2026-10-19.*

Flare drew the tidiest phone of the three, with the header, five days and icons, but its temperatures carry a degree sign and no C. Sunburst writes °C on every number and adds a tab bar. GPT-Image-2 fills the square with the screen and adds condition words, with °C after every figure. That makes two hits and a partial, and a client checking a mockup would catch the missing unit first.

## Verdict

Sunburst followed the most instructions. On these ten prompts at medium quality it finished with nine hits and one partial: it spelt every word, wrote every unit, kept the painting unsigned and cut out the glove. Its only slip was two cats sharing a gaze. GPT-Image-2 also scored nine hits and won the cats and the close-up, but it refused the transparent background outright. Flare finished with seven hits and three partials.

Flare has the most atmosphere: it lit the steam best and painted the loosest brushstrokes. Its slips, a mark like a signature, a shared gaze and a missing C, are the kind a reviewer catches and a client notices. Flare renders took 8 to 12 seconds each, Sunburst renders 11 to 18 and GPT-Image-2 renders 28 to 46. On this evidence, use Flare for drafts and moodboards where a person reviews every frame, and Sunburst for anything that goes straight to a client.

Price is where GPT-Image-2 loses outright on this test. OpenAI's [pricing page](https://developers.openai.com/api/docs/pricing) charges all three models $30 per million image output tokens. Here, though, each Flare and Sunburst render used 439 output tokens and each GPT-Image-2 render used 1,756, so the new models cost an estimated 1.3 cents a render against about 5.3 cents, a quarter as much. The guide warns that equal token rates do not mean equal cost per image, and here the difference is fourfold. The next thing to watch is the xhigh and max quality settings the 2.5 models add, which GPT-Image-2 does not offer.

## Scorecard

| Test | GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst | GPT-Image-2 |
|---|---|---|---|
| 1. Backlit steam at sunset | hit | hit, faintest steam | hit |
| 2. Exact text on a poster | hit | hit | hit, least minimal |
| 3. Five-step infographic | hit | hit | hit |
| 4. Cut-out on a transparent ground | hit | hit | miss, refused |
| 5. Impressionist style, no lettering | partial, mark like a signature | hit | hit |
| 6. Counting and placement | hit | hit | hit |
| 7. Two hands at work | hit | hit | hit |
| 8. Labelled real-world diagram | hit | hit | hit |
| 9. Three cats, three gazes | partial, two share a gaze | partial, two share a gaze | hit |
| 10. Interface with units | partial, no °C | hit | hit |
| Result | 7 hits, 3 partials | 9 hits, 1 partial | 9 hits, 1 miss |
| Estimated cost per render | about 1.3 cents | about 1.3 cents | about 5.3 cents |
| Time per render | 8 to 12 seconds | 11 to 18 seconds | 28 to 46 seconds |

## Frequently asked questions

### GPT-Image-2.5 Flare vs Sunburst: which should I use?

On our ten prompts at medium quality, Sunburst followed more instructions, with nine hits and one partial against Flare's seven hits and three partials. Flare returned faster, taking 8 to 12 seconds per render against Sunburst's 11 to 18.

### What does gpt-image-2.5-sunburst cost?

OpenAI lists [Sunburst](https://developers.openai.com/api/docs/models/gpt-image-2.5-sunburst) at $30 per million image output tokens, the same rate as Flare and GPT-Image-2. In our run, each 1024x1024 medium-quality Sunburst render cost an estimated 1.3 cents.

### What is the best OpenAI image model for text in images?

On these renders, Sunburst got every text demand right, including the °C on the weather screen that Flare dropped. All three models spelt the poster, the infographic and the planet labels correctly.

### Does GPT-Image-2 support transparent backgrounds?

No. Our request for a transparent background on GPT-Image-2 was refused with an error saying the model does not support it. OpenAI documents transparent PNG and WebP output for both GPT-Image-2.5 models.