# GPT Image 2.5 Sunburst — /images/generations

Image generation with GPT Image 2.5 Sunburst, the style branch of the next-generation OpenAI image model with heavier grain and brushwork, for pictures that need texture and physical presence. Same capabilities as Flare, as async jobs; 1K / 2K / 4K tiers and any aspect ratio, billed per image, nothing charged on failure.

**Endpoint:** `POST https://nezhagate.com/v1/images/generations`

## Quick start

Three steps: submit a job, poll it, take the image link. Replace YOUR_API_KEY with your key and the commands below run as they are.

**Step 1 · Submit (returns a job id at once, HTTP 202)**

```
curl https://nezhagate.com/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-sunburst", "prompt": "An oil painting of a harbour at dusk, visible brushstrokes, warm palette", "size": "16:9", "resolution": "2K"}'
```

Response (HTTP 202)

```
{
  "id": "img_7f3c9a1e2b4d6f8a0c1e3b5d",
  "object": "image.generation.job",
  "status": "queued",
  "created": 1791281400,
  "model": "gpt-image-2.5-sunburst"
}
```

**Step 2 · Poll every 2–3 seconds until `status` is `succeeded` or `failed`**

```
curl https://nezhagate.com/v1/images/jobs/img_7f3c9a1e2b4d6f8a0c1e3b5d \
  -H "Authorization: Bearer YOUR_API_KEY"
```

Response when it succeeds

```
{
  "id": "img_7f3c9a1e2b4d6f8a0c1e3b5d",
  "object": "image.generation.job",
  "status": "succeeded",
  "created": 1791281400,
  "model": "gpt-image-2.5-sunburst",
  "data": [
    {"url": "https://img.nezhagate.com/i/2be8cd16d37f3a4b9c0e1f2a3b4c5d6e.png"}
  ],
  "usage": {"images": 1, "resolution": "2K", "model": "gpt-image-2.5-sunburst"}
}
```

**Step 3 · `data[0].url` is the image (PNG). It is kept for 15 days; download it to your own storage to keep it longer.**

**Complete example (Python: submit, wait, save to disk)**

```
import time
import requests

BASE = "https://nezhagate.com/v1"
HEADERS = {"Authorization": "Bearer YOUR_API_KEY"}

# 1. submit
job = requests.post(f"{BASE}/images/generations", headers=HEADERS, json={
    "model": "gpt-image-2.5-sunburst",
    "prompt": "An oil painting of a harbour at dusk, visible brushstrokes, warm palette",
    "size": "16:9",
    "resolution": "2K",
}).json()
if "error" in job:
    raise SystemExit(job["error"]["message"])

# 2. poll every 3 seconds until it finishes
while job["status"] not in ("succeeded", "failed"):
    time.sleep(3)
    job = requests.get(f"{BASE}/images/jobs/{job['id']}", headers=HEADERS).json()

# 3. download the image
if job["status"] == "succeeded":
    url = job["data"][0]["url"]
    with open("gpt-image-2.5-sunburst.png", "wb") as f:
        f.write(requests.get(url).content)
    print("saved", url)
else:
    print("failed:", job["error"]["code"], job["error"]["message"])
```

## Authentication
```
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json
```

## Request parameters

The JSON body of `POST /v1/images/generations`:

| Parameter | Type | Required | Description |
| --- | --- | --- | --- |
| model | string | Yes | The model id: `gpt-image-2.5-sunburst`. |
| prompt | string | Yes | What to draw. For image-to-image, say what to change. |
| size | string | No | Aspect ratio such as `16:9` (recommended), or a pixel size such as `1536x1024`. Without it you get a 1:1 square. |
| resolution | string | No | Tier: `1K` (default), `2K` or `4K`. |
| quality | string | No | Detail level, at 1K only: `low`, `medium` (default) or `high`. 2K and 4K are always drawn at medium, and any other value counts as medium. It does not change the price; see Quality and native resolution below. |
| image / images | string / array | No | Reference images; sending any makes it image-to-image: one in `image`, several in `images` (up to 8). Each can be a public image URL, a data: URI or base64. |
| n | integer | No | Number of images. Best left out: for several images submit several jobs at once. With `n`, n images are held up front and the images actually delivered are billed. |
| callback_url | string | No | Where to push the result when the job ends; see Webhooks below. |

Other OpenAI parameters (`background`, `output_format`, `response_format`, `style` and the like) do not change the price and are not guaranteed to take effect; results are always PNG, returned as a `url`.

## Price

Billed per image by tier only: aspect ratio, text-to-image or image-to-image make no difference. The tier price is held when you submit and refunded in full if the job fails. 1 credit = $0.005.

| Tier | Credits per image | USD per image |
| --- | --- | --- |
| 1K | 1 | $0.005 |
| 2K | 6 | $0.03 |
| 4K | 10 | $0.05 |

## Aspect ratio, resolution and output size

Put the ratio in `size` and the tier in `resolution`. The image is delivered at exactly the size in the table: the tier sets the longest side and the other side follows the ratio. Ratios not in the table work the same way.

| Aspect ratio | 1K | 2K | 4K |
| --- | --- | --- | --- |
| 1:1 | 1024x1024 | 2048x2048 | 4096x4096 |
| 3:4 | 768x1024 | 1536x2048 | 3072x4096 |
| 2:3 | 683x1024 | 1365x2048 | 2731x4096 |
| 9:16 | 576x1024 | 1152x2048 | 2304x4096 |
| 4:3 | 1024x768 | 2048x1536 | 4096x3072 |
| 3:2 | 1024x683 | 2048x1365 | 4096x2731 |
| 16:9 | 1024x576 | 2048x1152 | 4096x2304 |

- At 1K the model composes the picture at the closest landscape, portrait or square shape, and the gateway then centre-crops and resizes it to the size in the table; the further a ratio is from 3:2, 2:3 or 1:1, the more of the edges is cut. The gateway adds a sentence about the ratio (in English) to your prompt so the subject stays whole and centred; keep anything that must survive away from the edges.

- The tiers that are drawn natively (see Quality and native resolution below) are drawn at your ratio, so nothing is cropped from the edges.

- `size` can also be a pixel size (e.g. `1536x1024`): the image is delivered at exactly that size (scaled down if the longest side is over 4096). The tier then comes from `resolution`, or without it from the longest side: up to 1536 is 1K, up to 2048 is 2K, larger is 4K. `resolution` can raise the tier but never lower it, and it does not change the delivered size, so do not pair a small size with a high tier.

## Quality and native resolution

- Without `quality` the image is drawn at OpenAI's **medium** detail level. At 1K you can choose `low`, `medium` or `high`, for text-to-image and image-to-image alike; **2K and 4K are always drawn at medium**. Any other value (`auto`, `xhigh`, `max` and so on) counts as medium and never causes an error. The price depends only on the resolution tier, never on `quality`.

- If `quality` holds `1K`, `2K` or `4K` and `resolution` is not given, it is read as the resolution tier (size and price follow that tier).

- **1K** is drawn at the model's native frame of about 1.57 megapixels (1254×1254 for a square), and the gateway resizes it to the size in the table above.

- **2K and 4K** are drawn natively at the ratio you ask for, up to 3840 px on a side and about 8.3 megapixels: 2K 1:1 comes back as 2048×2048, 4K 16:9 as 3840×2160 and 4K 1:1 as 2880×2880. Where the 4K size in the table is larger, the gateway scales the picture up to it (about 7% on each side for 16:9, more for squarer ratios); to receive the native pixels as they are, send a pixel `size` within those limits, such as `3840x2160`.

- Tiers that are drawn natively only run on their native lines: when those are busy the job waits in the queue, and if it still cannot be drawn it fails with a full refund — it is never replaced by an enlarged 1K image.

- A 1K job is occasionally finished on a backup line, where `quality` may not take effect.

## Image-to-image and reference images

- Send `image` (one picture) or `images` (several, up to 8; any beyond 8 are not used) and the job runs image-to-image; say in the prompt what to change. `POST /v1/images/edits` takes the same parameters.

- Both endpoints take a JSON body only, not a multipart file upload, so the OpenAI SDK's `images.edit()` cannot be used as is: pass links or base64 as in the example below.

- Each reference can be a public image URL (http or https), a data: URI or plain base64, up to 16MB each, in a common format such as JPEG, PNG or WebP.

- A URL has to download the image itself: not a web page, and not behind a login or hotlink protection. A reference we cannot read is refused at submit with a 400 that says which one and why; no job is created and nothing is charged.

- The output ratio comes from `size`, not from the reference. To keep the original ratio, set `size` to it.

- We host the references for you and keep them for 7 days.

**Image-to-image with two references**

```
curl https://nezhagate.com/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-sunburst", "prompt": "Keep the people, change the background to a beach at night with warm string lights", "images": ["https://example.com/people.jpg", "https://example.com/beach.jpg"], "size": "3:4", "resolution": "2K"}'
```

## Job status and fields

`GET /v1/images/jobs/{id}` reports these statuses; every field is described in [Async jobs](https://nezhagate.com/en/docs/guide/async-jobs#fields). Poll every 2–3 seconds, or use a webhook.

| Status | Meaning |
| --- | --- |
| queued | Waiting in line. Carries `queue_position` (jobs ahead of yours) and `eta_seconds` (expected wait). |
| processing | Rendering. |
| succeeded | Done. The result is in `data[0].url`, the billing detail in `usage`. |
| failed | Failed. The reason is in `error`; the credits held for the job are refunded in full. |

One account runs up to 50 jobs at a time; further jobs wait in submission order, and a job that has not started after 15 minutes fails and is refunded in full. Job records are kept for 3 days and image links for 15 days.

## How long it takes

From submit to result, 1K usually takes 20–60 seconds; 2K and 4K are drawn natively and usually take 30–120 seconds. Image-to-image and several references take longer, and a job the gateway retries on another line can take a few minutes.

Real calls over the last 7 days: median 26 s, 90% done within 44 s.

## Errors and refunds

An error at submit comes back at once as HTTP 4xx / 5xx: no job is created and nothing is charged. A failure after the job was accepted is reported by the job endpoint, and the hold is refunded in full. Every code is listed under [Error codes](https://nezhagate.com/en/docs/guide/errors).

### Errors at submit

| HTTP | code | Cause | What to do |
| --- | --- | --- | --- |
| 400 | missing_prompt | `/v1/images/generations` has no `prompt`. | Add a `prompt`. |
| 400 | invalid_request | A reference image cannot be read (dead link, not an image, over 16MB...; `message` says which one and why), `/v1/images/edits` has no reference image, the body is not a JSON object, or `n` is not an integer. | Fix what `message` says and submit again. |
| 400 | model_not_found | The model id is misspelled, `model` is missing, or the model cannot be used on this endpoint (a chat model on the image endpoint, for example). | Use a model id from `GET /v1/models` or the [API reference](https://nezhagate.com/en/docs/api). |
| 401 | invalid_api_key | The key does not exist, was deleted or is disabled, or the account is disabled. | Check the key on the API Keys page of the console. |
| 402 | insufficient_quota | Not enough balance to hold this request. The gateway holds the highest possible cost of the request first, so a burst of concurrent calls can trigger it too. | After a burst, wait a few seconds and retry; otherwise top up. |
| 429 | rate_limit_exceeded | Over the requests-per-minute limit set on this key (counted over the last 60 seconds). No `Retry-After` header. | Wait a few seconds and retry, or raise the key's limit in the console. |
| 503 | model_maintenance | The model is under maintenance and takes no requests; no other model answers in its place meanwhile. | Retry when maintenance ends; see [Status](https://nezhagate.com/en/status). |
| 503 | service_unavailable | A temporary failure on our side, such as storing a reference file; `/v1/responses` also answers 503, with `Retry-After`, when no line is available for a moment. | Retry shortly. |

### Failed jobs (`status` `failed`)

| code | When you see it | What to do |
| --- | --- | --- |
| moderation_blocked | Image: the model provider's safety system refused the prompt or a reference image (the code may also be `content_policy_violation`). | Rewrite the prompt or replace the reference; resubmitting it unchanged is usually refused again. |
| upstream_error | Any other failure: every line was tried, the job waited 15 minutes without starting, the upstream refused without a reason, and so on; `message` says which. | Submitting again usually works. |

## Webhooks instead of polling

Add `callback_url` when you submit and the moment the job ends, success or failure, the gateway pushes exactly what the poll endpoint returns to that URL, signed and retried on failure. You can also set an account-wide address under Settings → Webhook in the console. [Webhook docs →](https://nezhagate.com/en/docs/webhooks)

```
curl https://nezhagate.com/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.5-sunburst", "prompt": "An oil painting of a harbour at dusk, visible brushstrokes, warm palette", "size": "1:1", "callback_url": "https://your-domain.com/webhook"}'
```