GPT-Image-2.5 is split into two models:
Both accept text and image inputs, support the quality levels auto, low, medium, high, xhigh, and max, and work through the Images API for generation and editing.
For an existing OpenAI-compatible integration, the migration is small: change the base URL, credentials, and model ID. I would start with Flare at medium, measure latency and accepted-image cost, and route demanding edits or premium assets to Sunburst.
OpenAI introduced ChatGPT Images 2.5 and these two API models on September 8, 2026. Flare is positioned as the default choice for most applications; Sunburst is the more capable option for complex generation and editing. Both models are currently marked preliminary, and current Arena results favor Sunburst for generation and editing.
When using the OpenAI-compatible gateway, the relevant configuration is:
| Setting | Value |
|---|---|
| Base URL | https://api.cometapi.com/v1 |
| Generation route | POST /images/generations |
| Editing route | POST /images/edits |
| Authentication | Authorization: Bearer $COMETAPI_KEY |
Provider pricing and availability can change, so I verify model IDs, endpoint behavior, and billing in the dashboard before deploying.
Create a server-side token and keep it out of browser code, repositories, logs, screenshots, and client applications:
export COMETAPI_KEY="your-cometapi-key"
PowerShell:
$env:COMETAPI_KEY="your-cometapi-key"
For a first request, I usually use Flare with a controlled quality setting:
curl "https://api.cometapi.com/v1/images/generations" \
-H "Authorization: Bearer $COMETAPI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2.5-flare",
"prompt": "Premium product photograph of a matte black wireless speaker on a light concrete pedestal, soft window light, realistic material texture, clean editorial composition, no text",
"size": "1536x1024",
"quality": "medium",
"output_format": "png"
}'
The response normally contains generated image data in data[].b64_json, rather than a permanent image URL:
{
"data": [
{
"b64_json": ""
}
],
"usage": {
"input_tokens": 32,
"output_tokens": 1372,
"total_tokens": 1404
}
}
Decode the Base64 value and store the resulting bytes. The Base64 string should not become the final asset in your storage system.
import base64
import os
import requests
response = requests.post(
"https://api.cometapi.com/v1/images/generations",
headers={
"Authorization": f"Bearer {os.environ['COMETAPI_KEY']}",
},
json={
"model": "gpt-image-2.5-flare",
"prompt": (
"A clean isometric illustration of a solar-powered research lab, "
"white background, precise geometry, no labels or watermarks"
),
"size": "1536x1024",
"quality": "high",
"output_format": "png",
},
timeout=180,
)
response.raise_for_status()
payload = response.json()
image_b64 = payload["data"][0]["b64_json"]
with open("research-lab.png", "wb") as file:
file.write(base64.b64decode(image_b64))
An existing OpenAI SDK integration can use the same client pattern by changing base_url, the API key, and the model ID.
Use /images/edits when the input image must be preserved while only a defined region or attribute changes. Put preservation requirements before the requested modification:
curl https://api.cometapi.com/v1/images/edits \
-H "Authorization: Bearer $COMETAPI_KEY" \
-F "model=gpt-image-2.5-sunburst" \
-F "image[]=@product.png" \
-F "prompt=Preserve the product shape, label, and camera angle. Replace only the background with a warm studio gradient. Add no new text." \
-F "quality=high" \
-F "output_format=png"
For localized edits, a mask indicates where changes are allowed. Transparent pixels identify the editable region; the remaining area should be preserved. The mask must match the source image’s size and format, include an alpha channel, and stay within the API’s file-size limit. With multiple input images, the mask applies to the first image.
A mask is guidance, not a guaranteed pixel-perfect selection. I reinforce it in the prompt:
Change only the transparent region. Preserve all other pixels, text, and geometry.
Give every reference image a stable semantic role. I generally use subject first, style second, then background or layout. The prompt should specify which attributes may transfer from each image.
The Responses API is useful when image generation is one step inside a larger conversational or agent workflow:
import base64
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["COMETAPI_KEY"],
base_url="https://api.cometapi.com/v1",
)
response = client.responses.create(
model="gpt-6-astra",
input=[{
"role": "user",
"content": [
{
"type": "input_text",
"text": (
"Create a campaign image. Use image 1 only for the product "
"shape and colors; image 2 only for lighting and visual style; "
"image 3 only for the background composition. Preserve the "
"product logo exactly and add no other text."
),
},
{
"type": "input_image",
"image_url": "https://example.com/product.png",
},
{
"type": "input_image",
"image_url": "https://example.com/style.png",
},
{
"type": "input_image",
"image_url": "https://example.com/background.png",
},
],
}],
tools=[{
"type": "image_generation",
"model": "gpt-image-2.5-sunburst",
}],
)
for item in response.output:
if item.type == "image_generation_call":
with open("campaign.png", "wb") as file:
file.write(base64.b64decode(item.result))
The top-level Responses model is a supported mainline model; GPT-Image-2.5 is selected inside the image-generation tool. If the gateway does not expose that model or tool schema, use the currently documented equivalent.
Do not rely solely on upload order. Explicitly say “image 1 is the subject,” “image 2 is the style reference,” and so on. Also state which details must not be copied, such as faces, logos, text, or layout.
| Requirement | Image API | Responses API |
|---|---|---|
| One-shot generation or editing | Best fit | Usually unnecessary |
| Conversational or agentic flow | Limited | Best fit |
| Direct model selection | Set image model directly | Mainline model plus image tool |
| Multiple semantic references | Supported for edits, depending on route | Natural fit |
| Iterative turns and tool calls | Application-managed | Built in |
| Streaming previews | Supported | Supported |
My default is the Image API. I move to Responses when the workflow needs conversation state, several reference images with explicit roles, or other tools around image generation.
| Parameter | Purpose | Starting point |
|---|---|---|
quality |
Compute and detail level | medium during development |
size |
Resolution and aspect ratio | 1024x1024 or1536x1024 |
output_format |
PNG, JPEG, or WebP | PNG for fidelity |
background |
Opaque or transparent output | Transparent only when needed |
output_compression |
JPEG/WebP compression | Tune for delivery |
n |
Number of returned images | 1 |
prompt |
Visual requirements and constraints | Specify layout explicitly |
The supported quality ladder is:
auto low medium high xhigh max
auto lets the model choose. I prefer explicitly setting medium for controlled comparisons.
A practical deployment split:
low or medium: drafts, previews, and high-volume experimentation high: approved production assets xhigh or max: demanding final renders where the gain is measurable
The common presets are:
1024x1024
1536x1024
1024x1536
The 2.5 models also support arbitrary valid dimensions, useful for banners, product pages, mobile creatives, and other non-square assets. The current OpenAI specification allows up to 3840 pixels per edge within its pixel-count and aspect-ratio limits.
Use a format with alpha support:
{
"background": "transparent",
"output_format": "png"
}
WebP is also suitable. JPEG cannot represent transparent output. This mode is useful for product cut-outs, icons, stickers, UI assets, and compositing pipelines.
The model specification does not list generic model-level streaming, but the Images API and Responses API support image-generation streaming with partial_images. These are progressive previews, not token-by-token text output.
The Images API accepts partial_images values from 0 to 3. Each partial image adds 100 output tokens. A value of 3 does not guarantee three previews: if generation finishes quickly, fewer may arrive.
Set the value to 0 when previews do not improve the user experience.
I separate the creative goal from the constraints.
Specify the subject, framing, camera angle, depth, background, and object positions:
Three-quarter product view, centered, generous negative space on the right, eye-level camera, 50 mm lens look.
Describe direction, softness, contrast, color temperature, and material behavior:
Large softbox from the upper left, subtle rim light, realistic brushed aluminum, controlled reflections.
Quote required text and define its placement and typography:
Place the exact headline “BUILD WITH CLARITY” at the top center in bold uppercase sans serif. Preserve spelling exactly. Add no other words, letters, labels, or watermarks.
For edits, name everything that cannot change: identity, pose, product geometry, logo, label text, proportions, camera angle, and background.
Target likely failure modes rather than adding generic quality language:
No extra fingers, no duplicated products, no warped logo, no misspelled text, no border, no watermark.
| Workload | Flare | Sunburst |
|---|---|---|
| Interactive application | Recommended | Selective use |
| Rapid prompt iteration | Recommended | Usually unnecessary |
| High-volume generation | Recommended | Depends on acceptance rate |
| Product/reference editing | Good | Recommended |
| Complex final composition | Good | Recommended |
| Maximum editing control | Good | Recommended |
| Latency-sensitive UI | Recommended | Less suitable |
| Premium final asset | Test first | Recommended when the gain is measurable |
I would not choose one model permanently for every request. A sensible architecture routes routine traffic to Flare and sends difficult revisions or high-value final outputs to Sunburst.
At the time of verification, both models list the same token prices:
Actual cost depends on tokens used, not merely request count.
The gateway currently advertises a 20% discount for GPT-Image-2.5 Flare. I treat the dashboard and invoice as authoritative because gateway pricing can change.
Spend is also affected by:
The useful metric is accepted-image cost:
Accepted-image cost = total generation spend ÷ approved outputs
For example, 10 attempts at $0.18 cost $1.80. If six pass review, the accepted-image cost is $0.30. If better prompting reduces the run to eight attempts with six accepted images, it falls to $0.24.
If the existing application uses GPT Image 2, hold prompts, references, dimensions, and output format constant and change only the model during the comparison:
model = "gpt-image-2"
model = "gpt-image-2.5-flare"
model = "gpt-image-2.5-sunburst"
Then evaluate:
A model-ID swap is not enough for a production migration. Use a fixed evaluation set so the model is the variable being tested.
The surrounding service should remain boring:
400 responses differently from transient 429 and 5xx failures.
Do not retry every error. A malformed request will remain malformed, and retrying an authentication failure only creates more failed traffic.
| Error | Likely cause | Response |
|---|---|---|
401 Unauthorized |
Missing or invalid key | Check COMETAPI_KEY and the Bearer header |
400 Bad Request |
Invalid model, size, format, or parameter | Remove optional fields and test a minimal request |
429 Too Many Requests |
Concurrency or account limit | Retry with exponential backoff and jitter |
Repeated 5xx |
Temporary upstream issue | Retry a limited number of times |
| Base64 appears as text | b64_json was not decoded |
Decode and save the bytes |
| Transparent output fails | Incompatible format | Use PNG or WebP |
| Edit changes too much | Weak preservation constraints | State exactly what must remain unchanged |
| Unexpected cost increase | Higher quality, resolution, or retries | Log usage and calculate accepted-image cost |
| Tier | TPM | IPM |
|---|---|---|
| Tier 1 | 100K | 5 |
| Tier 2 | 250K | 20 |
| Tier 3 | 800K | 50 |
| Tier 4 | 3M | 150 |
| Tier 5 | 8M | 250 |
Flare is the practical default for fast, everyday generation. Sunburst is the better fit when preserving references, making localized edits, or producing a high-value final composition matters more than latency.
I would begin with /v1/images/generations, Flare, one representative prompt set, and an explicit quality level. Add /v1/images/edits and Sunburst after measuring the cases where Flare fails review.
The important production metric is not maximum quality in isolation. Measure latency, output-token usage, edit fidelity, acceptance rate, and cost per approved image on the workload the application actually serves.
Yes. GPT Image 2.5 Flare and GPT Image 2.5 Sunburst are available through the OpenAI-compatible gateway described above.
Yes. Both accept image inputs and support image editing. Use the edits route when an existing asset must be modified.
Start with Flare for most generation workloads. Use Sunburst when reference preservation, complex composition, or editing precision materially affects acceptance.
Yes. Set "background": "transparent" and use PNG or WebP. JPEG is not suitable for alpha transparency.
Yes. Instantiate the standard client with:
client = OpenAI(
api_key=os.environ["COMETAPI_KEY"],
base_url="https://api.cometapi.com/v1",
)
Then select either gpt-image-2.5-flare or gpt-image-2.5-sunburst as appropriate.