GPT-Image 2.5 Sunburst API
AI ImageGPT Image 2.5 Sunburst — Text-to-image generation and reference editing, priced per image in 1K, 2K and 4K tiers. Quality does not change the price.
- Per successful call$0.025–$0.055
- OperationsText to image
- Image edit
- View tool pricing and billing rules
Your result will appear here once you run the model.
01
Where it fits
GPT Image 2.5 Sunburst — Text-to-image generation and reference editing, priced per image in 1K, 2K and 4K tiers. Quality does not change the price.
Use Sunburst for reference-image edits and final assets: it is OpenAI's editing-precision variant, and our measurements show the same cost and behaviour as Flare at the same size and quality.
02
What it does
Text to image
gpt-image-2.5-sunburst
Turn a text prompt into a new image.
Image edit
gpt-image-2.5-sunburst
Change an existing image from a prompt and reference input.
03
In depth
GPT-Image 2.5 Sunburst is OpenAI's image generation and editing model on OmniaKey, the second of the two 2.5 ids beside GPT-Image 2.5 Flare. OpenAI describes Sunburst as the pair's choice for editing precision, and the catalog gives both ids the same capabilities and the same per-call tiers. When a picture carries words a reader will check — a menu, a price column, a label — this is the id the maker points at.
Sunburst is at its best where a layout has to survive contact with a reader. In our four-prompt comparison its two menu runs set every dish and price correctly, kept CAFE in capitals with its accent and used a right-aligned column, and it placed exactly five marbles in the left jar of the counting test. It lost the product-photo case, where the shoe ended up on an ankle in black jogger trousers, and the room edit dropped a bookshelf at the far left.
Two details from our runs are specific to this id. The extra quality levels OpenAI publishes did not add anything here: xhigh and max both rendered as medium with the same image-token count, and the 2.5 ids are absent from OpenAI's Batch table. Money is counted per image by output size, so a larger canvas moves the amount while a higher quality word does not. Read every word of the result before you ship it.
Menus and price columns
State the column rule in words — right-aligned prices, one dish per line — and Sunburst holds the grid, which is what our two menu runs did with every item, accent marks included. Ask for the accents explicitly, then proofread the last line, where our first run slipped in an extra cup.
Edits with a keep list
Write the prompt as two lists: what changes and what stays. Our living-room task asked for a new sofa and a plant; Sunburst kept the pictures, the books, the rug and even the white spine lettering, and set the plant where the bookshelf had stood. Naming each large object that must remain is the difference between a retouch and a rebuild.
Exact counts in a frame
When a frame has to show a set number of things, Sunburst came closest in our tests: five marbles, three lemons, an empty jar between them. Put each count beside its object in the prompt, line the objects up in one row, and ask for nothing to be hidden behind anything else.
04
Hands-on notes
In our cafe-menu test, GPT Image 2.5 Sunburst spelled every word and price right in both runs, including CAFE with an accent and the en dash in the opening hours.
Read the testIn our counting test, GPT Image 2.5 Sunburst put exactly five red marbles in the left jar and left the middle jar empty with its numbered tag.
Read the testIn our sneaker product photo, GPT Image 2.5 Sunburst added an ankle and black jogger trousers, turning the product shot into a lifestyle frame.
Read the testIn our living-room edit, GPT Image 2.5 Sunburst kept every listed item and the spine lettering but dropped the bookshelf and put the plant in its place.
Read the testOur 17 September 2026 measurements showed the xhigh and max settings on GPT Image 2.5 Sunburst rendering as medium with identical output image tokens.
Read the test
05
Limits to plan around
- The xhigh and max quality levels OpenAI publishes behaved like medium in our 17 September run, with identical image-token counts, so a request needs only low, medium or high.
- A product brief can slide into a lifestyle frame: our shoe sat on an ankle in black jogger trousers, so add a no-people clause and frame the object alone.
- An edit may remove a large object to make room. Our living-room run lost the bookshelf at the far left, so name every big item that must survive and check the corners.
- Words still need one proofing pass: the first menu run added a coffee cup at the edge of the card, which the prompt never asked for.
06
Compare nearby models
| Model | Best for | Operations | Listed price |
|---|---|---|---|
| GPT-Image 2.5 SunburstThis page gpt-image-2.5-sunburst | Menus, price columns and counted objects, where both menu runs and the marble count held. | Text to image · Image edit | $0.025–$0.055 per call |
| GPT-Image 2.5 Flare gpt-image-2.5-flare | Product photography and everyday frames, where props and reflections matter more than a price column. | Text to image · Image edit | $0.025–$0.055 per call |
| GPT-Image 2 gpt-image-2 | Bulk drafting on the older id, with Batch listed and custom pixel sizes accepted. | Text to image · Image edit | $0.02–$0.05 per call |
07
Sources and verification
Guide updated 2026-10-27.
08
Call it
Media models run asynchronously. Create a task, poll it until it finishes, then download the signed asset URL.
# 1. Create the job. 202 Accepted; the body carries the task id,
# the pre-charged price, and a poll_url. Reuse one Idempotency-Key
# across retries so a timeout never starts a second paid job.
curl --fail-with-body https://api.omniakey.com/v1/media/tasks \
-H "Authorization: Bearer $OMNIAKEY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: gpt-image-2.5-sunburst-001" \
-d '{
"model": "gpt-image-2.5-sunburst",
"type": "image",
"operation": "generate",
"input": {"prompt": "a paper boat on a quiet lake at sunset"},
"parameters": {"size": "1024x1024", "quality": "medium"}
}'
# 2. Poll GET /v1/media/tasks/{id} until status is
# succeeded | failed | canceled. Progress carries percent and stage.
curl --fail-with-body https://api.omniakey.com/v1/media/tasks/TASK_ID \
-H "Authorization: Bearer $OMNIAKEY_API_KEY"
# 3. Download assets[0].url before expires_at, or poll again for a
# fresh signed URL. Failed and canceled jobs are refunded.
curl -sS -o out.png "$(jq -r '.assets[0].url' result.json)"09
Pricing
Image jobs are billed per successful call, not per token. Failed or canceled jobs are refunded automatically.
$0.025–$0.055per callSave 53%
- 1K
- $0.025per call
- 2K
- $0.035per call
- 4K
- $0.055per call
More OpenAI image models
Related articles
GPT Image 2.5 API: Generate and Edit Images in Python
One POST creates the task and one GET reads it back. Here is the request code, the sizes we ran on October 29, 2026, and one edit shown before and after.
Read articleGPT Image 2.5 vs Nano Banana 2.1: Same-Prompt Test and Price
Four prompts, three image models, one day: what each made and what it really cost.
Read articleSeedream 5.0 Pro vs GPT Image 2: Same-Prompt Test and Price
Six jobs, the same prompts, three models: what each made and what it cost.
Read articleGPT-Image 2.5 vs GPT-Image 2
On OmniaKey both generations bill per image by size tier, and quality does not change the price. The 2.5 ids add no working quality tier, and requests need an explicit WIDTHxHEIGHT.
Read articleFrequently asked questions
How do I call GPT-Image 2.5 Sunburst on OmniaKey?
POST to /v1/media/tasks with the model id, type "image", and operation "generate" or "edit". Reuse one Idempotency-Key across retries so a timeout never starts a second paid job.
What does GPT-Image 2.5 Sunburst cost?
GPT-Image 2.5 Sunburst bills $0.025–$0.055 per call. Creation pre-charges that amount from your prepaid balance, and failed or canceled jobs are refunded automatically.
Why does the API return a task id instead of the image?
Provider latency runs from seconds to about a minute, so OmniaKey normalizes every image route into one asynchronous lifecycle. Poll GET /v1/media/tasks/{id} until the status is succeeded, then read the signed asset URL.
Sunburst or Flare: which one should I pick?
Split by the brief. Sunburst handled every word and price in our menu runs, while Flare made the better product photograph and kept a room intact. The two ids share a listed tier and take the same parameters, so one test script can cover both and your own frames decide.
Does Sunburst get menu text right?
In both of our runs, yes: every dish, price and the CAFE capitals came back as written, in a right-aligned column. The first run still added a cup at the edge of the card, so treat spelling and layout as settled and props as the open risk. Read the whole frame at full size.
Why do xhigh and max look the same as medium?
On this model they are the same: our 17 September probe found both settings produced medium quality with an identical image-token count, and the 25 October probe found the same for max. Send low, medium or high, and use the size to control detail instead.
How is GPT-Image 2.5 Sunburst charged?
Per image, at the rate in the price table on the model page, which steps by output size: the small frame is the lightest of the three, the middle one next, the largest the heaviest, and the quality word never moves the amount. A failed task returns its reserved credit.
Can Sunburst handle product photography?
For a single object on a built set we would start with Flare, which won that case in our comparison. Sunburst is the safer choice once the frame needs a caption, a price or a label. If you only adopt one id, run your own product brief twice through it before deciding.