OpenAI's image model in two tiers, Flare and Sunburst: legible text across scripts, edits that change only what you name, transparent PNGs and up to 4K. $0.05 to $0.07 per high-quality image at standard sizes.
GPT Image 2.5 is OpenAI's image generation and editing model, released on 8 September 2026 in two tiers: Flare for fast everyday work and Sunburst for precise edits. It renders legible text across scripts, changes only what a prompt names, and outputs transparent backgrounds and sizes up to 3840x2160. On Segmind both tiers run through one API key, $0.05 to $0.07 per high-quality image at standard sizes (as of 31 Oct 2026).
In-image text
GPT Image 2.5 spells what you put in quotes, in Latin and non-Latin scripts, from a headline down to an ingredient line. Every image here came back in one pass, untouched, like the Hindi and English signs in the opening image.
A bright red enamel sign board on the wooden front of a tiny ramen shop in Tokyo at night, lit by a paper lantern. White hand-lettered text: at the top in Japanese '本日のラーメン', below it in English 'TODAY'S RAMEN', then a price list with dotted leaders: '醤油 Shoyu ¥980', '味噌 Miso ¥1,050', '担々麺 Tantanmen ¥1,180'. At the bottom a small line: 'Open until 2 a.m.'. A noren curtain edge and steam from the doorway at the side. Photographed straight on, 50mm, warm lantern light, slight paint wear on the enamel.
This run: $0.052, rendered in 33 s
A clean editorial infographic poster titled 'POUR-OVER IN 4 STEPS' on an off-white background. Four numbered panels with simple line illustrations: '1. Rinse the filter' with a kettle over a paper filter, '2. Add 20 g of coffee' with a scale reading '20.0 g', '3. Bloom with 40 ml, wait 30 s' with a small timer, '4. Pour to 320 ml by 2:30' with a carafe. A footer bar reads 'Ratio 1:16 · Water 94 °C · Medium-fine grind'. Muted terracotta and charcoal palette, generous margins, a modern geometric sans-serif.
This run: $0.052, rendered in 21 s
A photoreal product photograph of a frosted glass bottle of artisanal hot sauce standing on a weathered wooden market stall, with three fresh red habanero peppers and a halved lime beside it. The wraparound kraft-paper label reads, in bold condensed serif lettering: 'EMBER & ASH' across the top, 'Smoked Habanero Hot Sauce' as the tagline beneath it, a small legible line 'habanero · smoked paprika · lime · sea salt', and '148 ml' in the lower corner. Condensation beads on the glass, a red wax-sealed cork, visible paper grain on the label, bright even daylight.
This run: $0.052, rendered in 20 s
Put each piece of text in quotes, exactly as it should read: 'SMOKED HABANERO HOT SAUCE', '148 ml'.
Name the position and the type: a title across the top in a tall serif, a price list with dotted leaders.
Low and medium soften small lettering. Use them for layout drafts, then render the final at high as a PNG.
OpenAI's guide notes that its image models can still struggle with precise text placement and clarity, so proofread anything you will print or publish.
Image editing
Give GPT Image 2.5 a photo and one instruction. It rewrites the label, recolours the product or fills the masked wall, and leaves the rest of the frame where it was. Drag across each image to compare.
Using the reference photo, change only the label: replace the word 'STRAWBERRY' with 'BLACKCURRANT' and replace the three red strawberries with a cluster of dark purple blackcurrants on a leafy stem. Keep the glass jar, silver screw lid, wooden shelf, grey background, lighting, and the label's shape, position and layout exactly the same.
This run: $0.082, rendered in 33 s
Recolor only the scooter's body panels to matte mustard yellow. Keep the chrome mirror, chrome headlight ring, handlebars, the white 'MU-42' number plate, the black seat, the cream plaster wall, the grey floor and the ground shadow exactly as they are.
This run: $0.067, rendered in 22 s
The mask: transparent where the print may go, opaque everywhere else.
Hang a single large framed botanical print on the bare wall above the sideboard: a dried fern study on cream paper in a slim oak frame with a wide mount. Match the room's soft morning light and cast a subtle shadow. Change nothing else.
This run: $0.067, rendered in 30 s
Send the photo in image_urls
Any public image URL switches GPT Image 2.5 to edit mode. Add more references to carry a subject or a style into the edit.
Name the change, then what stays
Say exactly what to change, then list what must not move: the label layout, the chrome, the shadow, the camera angle.
Add a mask for one region
mask_image_url takes a PNG with an alpha channel, the same size as the photo. Only its transparent area is repainted.
An edit also bills the input photo's tokens: these three cost $0.067 to $0.082, against about $0.05 to $0.07 for a new image of the same size.
Each step took the previous output as its only input. Her face, curls and earrings, the tram and the street held through a change of season, a new coat colour and a typeset headline: four runs, $0.254 in total.
StartStreet photograph of a Black woman in her thirties with short natural curls, wearing a mustard wool coat and a cream knit scarf, standing at a tram stop in Lisbon on an autumn morning, holding a paper coffee cup and looking down the street. A yellow vintage tram approaches in soft focus behind her, tiled facades and fallen leaves on the cobbles. Natural light, 50mm, shallow depth of field, she is placed right of centre with open sky above the rooftops on the left. The tram stop is a simple post with a blank sign; no logos, company names or brand names anywhere in the frame.
This run: $0.052, rendered in 34 s
Edit 1Change the season to deep winter: fresh snow on the cobbles, the tram tracks, the rooftops and the top of the tram stop sign, snow falling softly, a grey winter sky, bare branches instead of autumn leaves. Keep the woman exactly the same: her face, curls, earrings, mustard coat, cream scarf, bag, coffee cup, pose and gaze. Keep the yellow tram, the tiled facades, the blank stop sign and the camera angle the same.
This run: $0.068, rendered in 37 s
Edit 2Change only the colour of her wool coat from mustard to deep cobalt blue. Keep everything else exactly the same: her face, curls, earrings, cream scarf, bag, coffee cup, pose, the falling snow, the yellow tram, the street and the camera angle.
This run: $0.067, rendered in 35 s
Edit 3Add a minimal campaign headline in the open grey sky at the top left: 'WARM ROUTES' in a bold white condensed sans-serif, with 'Winter collection 2026' in smaller white type directly below it. Keep the photograph exactly the same: the woman, her cobalt coat, scarf, coffee cup, pose, the snow, the yellow tram, the street and the camera angle.
This run: $0.067, rendered in 34 s
Production files
Ask for a transparent background and GPT Image 2.5 returns a real alpha channel, ready for a sticker sheet, a storefront or an app. When a crop or a print needs the pixels, it renders up to 3840x2160.
On a light cardOn a brand colourA glossy die-cut vinyl sticker of a cheerful orange fox riding a red bicycle, thick white sticker border, with the words 'RIDE ON' in rounded bold letters on a small blue banner under the wheels. Flat illustration with soft highlights, isolated and centred on a transparent background.
This run: $0.066, rendered in 23 s
3840x2160, shown scaled down
Full-resolution cropA top-down photograph of a ceramicist's glazing table in soft north light. In the centre, a hand-thrown teapot in deep celadon glaze with a fine crackle pattern and a glaze drip near the spout. Beside it, a row of six glazed test tiles, each with a strip of masking tape hand-labelled in pencil: 'Celadon', 'Tenmoku', 'Shino', 'Ash', 'Chun', 'Oribe'. An open notebook page reads 'Cone 10 reduction · 12 Nov' in neat handwriting. Wide flat brushes, a natural sponge, a wooden rib and a coil of raw stoneware on a raw linen cloth. No logos or brand names anywhere. Extremely fine detail edge to edge: clay grain, glaze pooling, pencil texture.
This run: $0.126, rendered in 46 s
Two models, one family
Segmind serves GPT Image 2.5 as two models behind one API key, with the same parameters. At the same size and quality they cost the same per image; Sunburst spends longer on each one.
gpt-image-2.5-flare
OpenAI describes it as fast, high-quality everyday image generation. It edits too, with the same parameters as Sunburst.
Pick it for volume: product shots, social posts, mockups, and drafts at low quality for under a cent each.
gpt-image-2.5-sunburst
OpenAI calls it its most capable image model, for workflows where editing precision matters most.
Pick it for hero creative, multi-step edits and dense layouts, where a retry costs more than the extra wait.
A sunlit farmers' market stall piled with heirloom tomatoes, peppers and herbs. A smiling stallholder in her sixties with silver braids and a denim apron weighs tomatoes on an old brass balance scale with no markings. Handwritten chalkboard tags read 'HEIRLOOM TOMATOES $4/lb', 'SHISHITO PEPPERS $3', 'BASIL 2 for $5', and a painted wooden sign above the stall reads 'WILLOW CREEK FARM · EST. 1987'. No other signs or brand names. Late afternoon light, shallow depth of field, documentary photography.
# Generate: swap the slug for gpt-image-2.5-sunburst to change tier
curl -X POST "https://api.segmind.com/v1/gpt-image-2.5-flare" \
-H "x-api-key: $SEGMIND_API_KEY" \
-H "Content-Type: application/json" \
-o poster.png \
-d '{
"prompt": "A screen-printed travel poster titled THE COASTAL RAIL ROUTE...",
"size": "1536x1024",
"quality": "high"
}'
# Edit: send the photo in image_urls, optionally with an alpha mask
curl -X POST "https://api.segmind.com/v1/gpt-image-2.5-sunburst" \
-H "x-api-key: $SEGMIND_API_KEY" \
-H "Content-Type: application/json" \
-o edit.png \
-d '{
"prompt": "Change only the label to BLACKCURRANT. Keep everything else the same.",
"image_urls": ["https://example.com/jam-jar.jpg"],
"size": "1024x1024",
"quality": "high"
}'The response is the image file itself. Every parameter, with Python and JavaScript samples, is in the GPT Image 2.5 Flare API reference and the GPT Image 2.5 Sunburst API reference.
| Specification | Value |
|---|---|
| Models | GPT Image 2.5 Flare (gpt-image-2.5-flare), GPT Image 2.5 Sunburst (gpt-image-2.5-sunburst) |
| Inputs | A text prompt, optional reference images in image_urls, an optional mask_image_url |
| Size presets | 1024x1024, 1536x1024, 1024x1536, 1280x960, 960x1280, 1536x864, 864x1536, 2048x2048, 2048x1152, 1152x2048, 3840x2160, 2160x3840, or auto |
| Size limits | Sides in multiples of 16, longest side up to 3840, aspect 1:3 to 3:1, 655,360 to 8,294,400 pixels; above 2560x1440 is experimental per OpenAI |
| Quality | Low, medium, high (the default) or auto |
| Background | Opaque (the default), transparent or auto |
| Output format | PNG (the default), JPEG or WebP, with compression from 0 to 100 |
| Mask | A PNG with an alpha channel, the same size as the photo; it applies to the first reference image |
| Moderation | Auto (the default) or low |
| Response | The image file itself from /v1, or a request id to poll from /v2 for queued production jobs |
| Provenance | C2PA content credentials in the PNG; an invisible SynthID watermark, per OpenAI |
LMArena's blind human votes rank GPT Image 2.5 Sunburst first for both generating and editing images, with GPT Image 2.5 Flare second on both boards.
| Model | Text to image | Image edit |
|---|---|---|
| GPT Image 2.5 Sunburst | 1427 (#1) | 1517 (#1) |
| GPT Image 2.5 Flare | 1401 (#2) | 1482 (#2) |
| GPT Image 2 (medium) | 1381 (#3) | 1460 (#4) |
| MAI-Image-2.6 | 1329 (#4) | 1427 (#6) |
| Nano Banana 2.1 | 1325 (#5) | 1427 (#7) |
| Grok Imagine Image 2.0 | 1317 (#6) | 1439 (#5) |
| MAI-Image-2.7 Preview | Not listed | 1471 (#3) |
Sources: LMArena text-to-image leaderboard and image edit leaderboard. Artificial Analysis's text-to-image arena also ranks Sunburst first (1222) and Flare second (1207), tested at OpenAI's max quality setting, which Segmind does not offer (31 Oct 2026).
The same prompt at 1024x1024 and high quality, one run on each model on 31 Oct 2026. GPT Image 2 is billed at the same token prices as GPT Image 2.5, yet this image cost four times as much: GPT Image 2.5 spends far fewer output tokens on it.
Prompt: A ripe pear on a linen napkin, soft window light, still-life photograph.
| Feature | GPT Image 2.5 | GPT Image 2 | GPT Image 1.5 |
|---|---|---|---|
| Price per million tokens | $6.25 text in, $10 image in, $37.50 out | $6.25 text in, $10 image in, $37.50 out | $10 in, $40 out |
| Average wait, 30 days | Flare about 21 s, Sunburst about 39 s | About 27 s | About 31 s |
| Largest size | 3840x2160 | 3840x2160 | 1536x1024 |
| Transparent background | Yes | No (opaque or auto) | Yes |
| Editing | Same model, with an optional mask | Same model, with an optional mask | A separate model, GPT Image 1.5 Edit |
More on the previous generation: GPT Image 1.5 overview and the GPT Image 2 model page.
Mock up a label with every line legible, then rewrite the flavour, the weight or the language in an edit instead of a new shoot.
Generate the hero image, then change the season, the colourway or the headline in follow-up edits that keep the person and the set.
Recolour one product photo into every variant, and cut products out on transparent backgrounds for any page or marketplace.
Bilingual boards and event posters with prices, dates and names set in the image, in Latin and non-Latin scripts.
Numbered steps, labelled diagrams and simple charts for docs, decks and social posts, laid out from one prompt.
Transparent PNGs for sticker sheets, app icons, empty states and game assets, ready to drop onto any background.
Pay per token, with no subscription. Flare and Sunburst share one rate card, so the same request costs the same on either; size and quality decide the bill.
| Tokens | Rate |
|---|---|
| Text input | $6.25 per million tokens |
| Image input (edits) | $10 per million tokens |
| Output | $37.50 per million tokens |
| Request | Cost | Ran on |
|---|---|---|
| 1024x1024, low | $0.0075 | Flare |
| 1024x1024, medium | $0.017 | Flare |
| 1024x1024, high | $0.066 | Flare and Sunburst |
| 1536x1024 or 1024x1536, high | $0.052 | Flare and Sunburst |
| 2048x1152, high | $0.054 | Flare and Sunburst |
| 3840x2160, high | $0.126 | Flare and Sunburst |
| Edit at 1536x1024, high, one input photo | $0.067 | Flare and Sunburst |
| Edit at 1024x1024, high, one 2048x2048 photo | $0.082 | Sunburst |
GPT Image 2.5 Flare at 1024x1024, one prompt: A ripe pear on a linen napkin, soft window light, still-life photograph.
GPT Image 2.5 is OpenAI's image generation and editing model, released on 8 September 2026. It comes in two tiers, Flare for fast everyday work and Sunburst for precise edits, and renders legible text, transparent backgrounds and images up to 3840x2160. Segmind runs both GPT Image 2.5 tiers through one pay-as-you-go API.
GPT Image 2.5 Flare is the fast default, about 21 seconds per image on Segmind. GPT Image 2.5 Sunburst takes about 39 seconds and is OpenAI's choice for work where editing precision matters most. At the same size and quality both cost the same per image, so choose by speed against control (30-day averages, 31 Oct 2026).
GPT Image 2.5 is billed per token on Segmind: $6.25 per million text input tokens, $10 per million image input tokens and $37.50 per million output tokens. In measured runs a 1024x1024 image cost $0.0075 at low quality and $0.066 at high, and a 3840x2160 image cost $0.126 at high (31 Oct 2026).
Send the photo in image_urls, describe the change, then list what must stay the same; GPT Image 2.5 switches to edit mode whenever image_urls is set. To limit the change to one region, add mask_image_url: a PNG with an alpha channel, the same size as the photo, transparent where GPT Image 2.5 may paint.
Yes. Set background to transparent and output_format to png or webp, and GPT Image 2.5 returns an image with a real alpha channel, ready for stickers, logos, product cutouts and UI assets. GPT Image 2 on Segmind offers only opaque or auto backgrounds.
GPT Image 2.5 outputs up to 3840x2160 or 2160x3840. Each side must be a multiple of 16, the aspect ratio between 1:3 and 3:1, and the total between 655,360 and 8,294,400 pixels. OpenAI calls sizes above 2560x1440 experimental. A 4K image at high quality cost $0.126 on Segmind (31 Oct 2026).
Yes. GPT Image 2.5 renders non-Latin scripts: the Hindi shop signs and the Japanese ramen menu on this page each came out in one pass on Segmind. OpenAI notes its image models can still struggle with precise text placement, so put every line in quotes in the prompt, keep quality on high and proofread before you publish.
GPT Image 2.5 ranks above GPT Image 2 on LMArena: Sunburst 1427 and Flare 1401 against 1381 for text to image, and 1517 and 1482 against 1460 for image edit (October 2026). At the same token prices on Segmind, one 1024x1024 high-quality image cost $0.066 against $0.264, and GPT Image 2.5 adds transparent backgrounds (31 Oct 2026).
Yes. The PNG files GPT Image 2.5 returns on Segmind carry C2PA content credentials that mark them as AI generated, and OpenAI says its API also embeds an invisible SynthID watermark. Converting or re-saving a file can strip the C2PA metadata, so keep the original PNG where provenance matters.
Create a Segmind API key, then POST a prompt to https://api.segmind.com/v1/gpt-image-2.5-flare, or gpt-image-2.5-sunburst, with the key in the x-api-key header. GPT Image 2.5 returns the image file in the response. Each model page has curl, Python and JavaScript examples and every parameter.
One API key, both GPT Image 2.5 models, from under a cent for a low-quality draft.
Last updated