Connect to GPT Image 2.5 Flare in one click on RequestBeam. One API key connects AI and multiple data services, so you can focus on your business and explore new ideas.
0-1M input tokens
Text input
Official price
Our price Save 20%
Cached text input
Official price
Our price Save 20%
Image input
Official price
Our price Save 20%
Cached image input
Official price
Our price Save 20%
Image output
Official price
Our price Save 20%
Sign in to run with your account balance.
Example preview
{
"data": [
{
"url": "/examples/gpt-image-25-studio-product-gpt-image-2-5-flare.webp"
}
]
} Sign in to run with your account balance.
Example preview
{
"data": [
{
"url": "/examples/gpt-image-2-edit-background.webp"
}
]
} These before/after edits were made with GPT Image 2, not Flare. To try a prompt with Flare, host the pictured source at a public URL or adapt the bottle details to your own image. Results can differ.
Edit the supplied product photo. Replace only the kitchen setting with a seamless light gray studio backdrop and neutral tabletop. Keep the same teal bottle in the same position and scale; preserve its silver cap, black base ring, matte finish, proportions, and color. Match the light direction and retain a natural contact shadow. No text, logos, extra objects, or watermark.
Transform the supplied product photo into a hand-painted watercolor illustration on subtly textured white paper. Keep the bottle, plant, countertop, framing, and overall composition recognizable. Preserve its teal body, silver cap, black base ring, and upright proportions. No text, logos, or added objects.
Edit the supplied product photo. Change only the teal body of the bottle to a matte red finish. Keep the silver cap and black base ring, shape, scale, position, kitchen background, left-side lighting, and contact shadow the same. Do not add text, logos, or other objects.
We position GPT Image 2.5 Flare as an AI image model for everyday image generation, commercial asset production, and fast visual iteration. OpenAI released the GPT Image 2.5 series on September 8, 2026, with two versions: Flare and Sunburst. Flare focuses more on generation speed and cost efficiency, while Sunburst is better suited to work that requires greater detail and editing precision.
Compared with GPT Image 2, Flare's main value is not just generating beautiful images, but enabling users to complete multiple rounds of experimentation with less waiting. OpenAI positions it as the default choice for most applications and states that its image generation latency is approximately 50% lower than GPT Image 2, while also improving image detail, reference image preservation, and editing capabilities. OpenAI's official release announcement
We believe Flare is worth testing first for users who frequently need to generate advertising images, social media assets, product images, and product prototypes. However, we do not believe it surpasses GPT Image 2 on every image quality metric. If your output requirements are less demanding but speed matters more, Flare may be a good choice.
GPT Image 2.5 Flare can accept text and image inputs and generate new images. It can create visual content directly from written descriptions or modify existing photos, design drafts, and reference images.
We believe these two capabilities suit different workflows. If you simply want to generate a new product promotional image, you can directly describe the subject, composition, lighting, and style. If you already have an image you are mostly satisfied with and only need to replace the background, adjust a person's clothing, or change the text in the image, you can use the image editing endpoint instead of starting from scratch.
At the API level, Flare supports several quality settings, including low, medium, high, xhigh, max, and auto. Output dimensions are customizable, supporting images up to approximately 4K within the official pixel limits. It can also generate PNG or WebP assets with transparent backgrounds.
We recommend choosing quality and dimensions based on the final use rather than setting every task directly to max. For example, ordinary website thumbnails and social media previews can start at medium; for main product images that need to show extensive detail, increase the quality gradually.
We believe one important change in GPT Image 2.5 Flare is its greater attention to which parts of an image should not be modified.
In the past, AI image editing often presented a problem: you only wanted to change the color of clothing, but the person's face, background, or overall composition changed as well. For commercial design, this means previously approved assets may need to be reviewed again.
In GPT Image 2.5, OpenAI focused on improving local edits, subject preservation from reference images, and successive revisions. In the official release materials, Higgsfield AI's Head of Product, Axultan Alimkulov, specifically praised Flare's ability to preserve people, composition, and visual identity when modifying images. OpenAI early customer feedback
We think this capability is particularly valuable for product photography, virtual models, and brand assets. For example, e-commerce teams can try producing successive advertising assets with different backgrounds while keeping the product's shape, Logo, and label unchanged.
However, this does not mean it can keep everything perfectly unchanged at the pixel level. We still recommend explicitly telling the model what may be modified and what must be preserved, and inspecting the final image, especially faces, Logos, product packaging, and text on products.
Through RequestBeam, we want to help you understand more quickly how GPT Image 2.5 Flare performs in real business scenarios.
Simply sign in to RequestBeam, create an API Key, and add a small amount of credit to start trying the model's image generation capabilities through the online tool. During evaluation, you do not need to write API integration code first or purchase a fixed plan for a short-term test. It also reduces the hassle of maintaining accounts on different model platforms yourself.
We believe one of GPT Image 2.5 Flare's main advantages is reducing the time cost of creative trial and error.
For advertising and social media professionals, whether an image is ready to use often depends on many details, including the product angle, background color, a person's pose, and the overall style. Even with a clear creative direction, it is usually still necessary to generate multiple versions to choose from.
In the real API tests published by OneWave on September 8, 2026, Flare's median image generation time at 1024×1024 and high quality was 16.7 seconds, compared with 117.8 seconds for GPT Image 2. We believe a speed difference like this directly changes how users produce assets because trying a new idea involves significantly less waiting. OneWave hands-on test
For users who need to produce many advertising variants, we recommend using Flare during creative exploration, such as preparing several different background, camera, and style directions at once, then selecting versions worth refining further.
Of course, speed cannot replace quality review. We still recommend checking product descriptions, brand elements, and human details in advertising images. In particular, do not skip the final human review just because generation is faster.
GPT Image 2.5 Flare is suitable for generating product photography, lifestyle images, and photorealistic assets. OpenAI emphasizes the new series' improvements in natural lighting, surface textures, and reference fidelity, which we also consider among the more noticeable improvements for everyday users.
One of our customers shared that a portrait generated in a single attempt with Flare satisfied them without any further editing. The creator said the lighting and material rendering in the image were the most surprising aspects.
Of course, not every user achieves their ideal result with Flare. Based on our experience, we recommend describing all the details you need as clearly as possible when using it. For example, when asking Flare to generate product images for a coffee brand, we recommend supplying a packaging reference image and explicitly requiring it to preserve the product proportions, Logo, label, and brand colors, rather than only entering "generate a high-quality coffee advertising image."
From our customer feedback, we have found that image editing accounts for a greater share of everyday work than generation from scratch, with our current data showing a ratio close to 3:1.
For example, the composition and character design of an advertising image may already be complete, but the client wants a different clothing color or the product placed against a new background. The ideal result is to modify only the specified content while keeping other approved details consistent.
In an independent test on September 8, 2026, Genflick tried changing a character's mustard-yellow scarf to cobalt blue while requiring the face, glasses, pose, and background to remain unchanged. At the same canvas aspect ratio, both Flare's medium and high test samples completed the edit while preserving the other major visual elements. Genflick image editing test
However, we note that this preservation is visual similarity, not identical pixels outside the edited region.
For ordinary product images, poster revisions, and clothing changes, we therefore recommend trying Flare first. For brand Logos, complex product structures, or commercial assets that require strict preservation of the original image, Sunburst should be tested further, or traditional image editing tools should be used alongside the model for fine adjustments.
We believe multiple-reference-image capabilities are highly valuable for short-drama storyboards, comic characters, product combinations, and brand visual production.
Suppose you have prepared character reference sheets for two people and a background image for a fixed scene. Your next requirement might be to have these two characters perform a specified action in that scene while retaining their original clothing and character traits.
One of our animated-comic customers described a test scenario using three reference images in total: two characters + one scene. They requested an image of one character handing a cup to the other. In the demonstrated medium and high results, Flare preserved the main character traits, clothing, and environmental elements, and also arranged the characters' positions and actions as requested. This result leads us to believe Flare can already be used in pre-production for some sequential visual content.
Many earlier AI image models could generate attractive-looking posters, but spelling and number errors were common when images needed to include accurate text, dates, or product prices.
In actual footwear and apparel poster tests by our e-commerce customers, Flare also rendered the specified headlines, subheadings, and prices correctly. From this test, we see that GPT Image 2.5 Flare can already be used for practical design work in this area.
Beyond text, transparent backgrounds are another useful capability. The evaluation team at Manus praised Flare's improvements in transparent-background generation in OpenAI's official release materials.
We recommend using Flare for initial event poster drafts, product stickers, website illustrations, social media cards, and simple icons. For images requiring transparent backgrounds, you can use AI to develop a tool that checks the output file's Alpha channel to avoid possible errors.
In this section, we primarily refer to OpenAI's official materials and test results published by OneWave, Genflick, HummingBytes, IMG.LY, and Fuser.
Note that the resolution, quality settings, Prompts, and runtime environments are not identical across tests. We therefore retain important test conditions rather than combining results from different sources directly into a single model ranking.
In independent tests conducted by OneWave on September 8, 2026, three models used the same Prompt, a resolution of 1024×1024, and the high quality setting. Each model underwent 9 generation tests.
The results were as follows:
| Model | Median Generation Time | Cost per Image |
|---|---|---|
| GPT Image 1 | 30.3 seconds | $0.1669 |
| GPT Image 2 | 117.8 seconds | $0.2112 |
| GPT Image 2.5 Flare | 16.7 seconds | $0.0532 |
In this test set, Flare generated images approximately 7 times as fast as GPT Image 2, a very clear advantage. Another set of tests by HummingBytes observed a similar trend. Across its 19 high quality generations, Flare's average completion time was approximately 25.1 seconds, compared with 128.4 seconds for GPT Image 2. The absolute times differed between the two independent tests, but both showed Flare's clear advantage in image generation speed. OneWave test · HummingBytes test
If your business requires users to try different Prompts frequently, such as an AI image generation website, marketing asset tool, or e-commerce product image generator, we recommend testing Flare first.
However, for online services that need to promise a fixed response time, we recommend measuring P95 and P99 latency in addition to average generation speed, and implementing appropriate timeouts and retries. Do not treat the 16.7 seconds in these tests as a stable service response time.
Seconds · Lower is better
Independent test results, not a guaranteed response time. Do not combine with the separate HummingBytes test.
Source: OneWave · Zero-based scale
Seconds · Lower is better
Approximate means, not medians. This is a separate test with different conditions.
Source: HummingBytes · Zero-based scale
Another advantage of Flare is its generation cost.
According to OpenAI's official API pricing, GPT Image 2.5 Flare's standard processing prices are as follows:
| Token Type | Price per Million Tokens |
|---|---|
| Text input | $5.00 |
| Cached text input | $1.25 |
| Image input | $8.00 |
| Cached image input | $2.00 |
| Image output | $30.00 |
It is worth noting that Flare and GPT Image 2 have the same per-Token prices. The difference in actual image generation cost mainly comes from Token consumption during generation, rather than a reduction in the unit Token price. OpenAI Flare model pricing
In OneWave's test, with the same Prompt, resolution, and high quality, Flare cost $0.0532 per image, compared with $0.2112 for GPT Image 2. Flare's actual cost was approximately one quarter of the older version's.
We believe this is very important for SaaS products that need to generate images at scale. For example, if an AI marketing tool allows users to repeatedly modify Prompts, generate candidate images, and select a final version from multiple results, a lower cost per image directly affects the entire product's gross margin.
Before a formal integration, we recommend separately testing image quality and Token consumption at medium, high, and xhigh to find the lowest quality setting users can accept.
USD per image · Lower is better
Per-image costs from this test, not per-million-token prices or RequestBeam prices.
Source: OneWave · Zero-based scale
We consider text accuracy within images an important indicator of whether an AI image model can enter commercial design workflows.
OneWave used three models to generate a pricing comparison table containing 24 specified text items, including numbers, decimals, prices, and specific strings.
In this test, GPT Image 2 and GPT Image 2.5 Flare both correctly generated all 24 items, while GPT Image 1 correctly generated only 19.
In addition, some of our e-commerce users recently compared three tasks: an e-commerce Banner, a product detail image, and a product brochure. Flare correctly generated all 583 requested characters across the three test images.
We therefore believe these results show that Flare can already handle many types of text-bearing visual content, such as event posters, menus, product packaging concepts, and social media assets with explanatory text.
For general promotional posters, we recommend using Flare to produce the first draft. If your business scenario has strict text requirements, we instead recommend using Flare to generate the background and visual elements, then adding the final text with a programmatic layout tool, which is usually more reliable.
Correct text items · Higher is better
One task, not an overall accuracy rate. The separate customer test of 583 characters is not included.
Source: OneWave · Zero-based scale
In real commercial design, a Prompt often does more than describe a subject: it includes multiple requirements that must be met simultaneously.
OneWave designed a complex Prompt of approximately 1253 characters containing 8 conditions to verify, such as object counts, placement order, text positions, lighting direction, and elements that must not appear.
In this test:
| Model | Constraints Met |
|---|---|
| GPT Image 1 | 6/8 |
| GPT Image 2 | 7/8 |
| GPT Image 2.5 Flare | 8/8 |
| GPT Image 2.5 Sunburst | 8/8 |
These results show that, for this task, Flare not only correctly understood which objects to generate but also handled spatial relationships between objects and exclusion conditions. OneWave complex Prompt test
However, we believe this success should be distinguished from complete control over artistic style.
In feedback from one of our customers, their business workflow involved nine-panel storyboards. After switching from GPT Image 2 to Flare, the same Prompt correctly generated nine frames in a single attempt, preserving the main character traits and largely following the specified shot order. However, although the Prompt requested an illustration style, the actual result still leaned noticeably toward photorealistic cinematic imagery.
This shows that understanding scene structure does not mean the model will necessarily follow every visual style requirement strictly. Based on our experience, we recommend using structured Prompts that describe the subjects, positions, counts, text, and content that must remain unchanged separately. This tends to produce more stable image results than an unstructured text description.
If a project places particular emphasis on a consistent visual style, such as comics, brand illustrations, or animation storyboards, we recommend adding reference images to make the output more controllable.
Satisfied constraints · Higher is better
Task-specific results do not establish full control over artistic style.
Source: OneWave · Zero-based scale
OpenAI identifies precise image editing as one of GPT Image 2.5's main improvements, but we believe editing results and editing speed still need to be examined separately.
OneWave used the same source image in its editing tests and obtained the following results:
| Model | Median Editing Time | Cost per Edit |
|---|---|---|
| GPT Image 1 | 38.7 seconds | $0.1686 |
| GPT Image 2 | 15.3 seconds | $0.0143 |
| GPT Image 2.5 Flare | 17.0 seconds | $0.0143 |
In these tests, Flare was not faster than GPT Image 2; it was slightly slower, and the editing costs were identical.
On the other hand, Genflick's character scarf recoloring tests for Flare and Sunburst showed that both models could complete local color changes while preserving the main character traits and the original canvas aspect ratio. Genflick's transparent teapot icon test also verified that Flare can output a transparent PNG that genuinely contains an Alpha channel. OneWave editing performance test · Genflick editing and transparent-background tests
We believe Flare offers greater upgrade value if your business primarily generates images from scratch. However, if users mainly make successive edits to existing images, there is no need to switch immediately just because a new version has been released.
Our recommendation is to keep GPT Image 2 if it already meets your business needs, and consider expanding Flare's use only when it clearly reduces erroneous changes and manual rework.
Seconds · Lower is better
Flare was slightly slower than GPT Image 2 in this editing test.
Source: OneWave · Zero-based scale
USD per edit · Lower is better
Flare and GPT Image 2 had the same cost in this test. These are not RequestBeam prices.
Source: OneWave · Zero-based scale
We believe GPT Image 2.5 Flare's main advantage over GPT Image 2 is completing image generation tasks of similar quality with lower time and Token costs.
In OneWave's tests under the same conditions, Flare's median generation time was 16.7 seconds, compared with 117.8 seconds for GPT Image 2. The corresponding costs per image were $0.0532 and $0.2112.
Based on these figures, we are more inclined to choose Flare first for new projects that need frequent image generation.
However, in tests editing existing images, GPT Image 2 actually had a slightly lower median time, and both models had the same cost. This shows that Flare's speed advantage does not extend to every editing task. OneWave comparison test
Our recommendation is this: if you use GPT Image 2 to build a text-to-image product, prioritize testing a migration to Flare. If the product mainly provides photo modifications, background replacement, and local edits, first compare editing success rates and subject preservation before deciding whether to upgrade.
For new projects, we recommend putting Flare first on the default shortlist because it has more pronounced practical advantages in speed and cost.
GPT Image 2.5 Flare and Sunburst belong to the same model family, but they do not target exactly the same work.
OpenAI's official definitions are clear: Flare is a smaller model that prioritizes speed, while Sunburst is a base model that places greater emphasis on image quality and editing precision.
Across HummingBytes' 19 high quality generation tests, Flare's average time was approximately 25.1 seconds, compared with 36.3 seconds for Sunburst. Flare was noticeably faster in this test set. HummingBytes hands-on comparison
In OneWave's complex Prompt test, both Flare and Sunburst met all 8 constraints. Some of Genflick's simple poster and local recoloring tests also found no clear advantage in task completion rates from moving up to Sunburst.
We believe that when quality requirements are already met, using Sunburst may only add waiting time without delivering a sufficiently clear benefit.
We recommend a two-stage model strategy: first use Flare to generate different creative versions quickly, then try Sunburst for high-precision revisions after the user settles on a composition and style.
For example, an advertising design tool could use Flare to generate ten concept images, then use Sunburst to refine the details of the final selected image. This makes it easier to balance efficiency and quality.
Google's Nano Banana 2.1 is one of the image models we believe is worth evaluating alongside Flare.
On October 6, 2026, Fuser conducted a set of hands-on tests containing five tasks to compare Nano Banana 2.1 and GPT Image 2.5 Flare.
In these tests, Nano Banana 2.1 performed better in product photography, handling the specified backgrounds and shadows more accurately. Flare had advantages in character consistency, some photorealistic scenes, and product background editing. Both correctly generated the requested content in the specified poster text test. Fuser hands-on test of both models
We believe these results are useful because they show how closely an image model's strengths depend on the specific task.
If your product needs to preserve a virtual character's facial traits and clothing details across multiple new scenes, we recommend testing Flare first.
If your business mainly involves product photography, product presentations, and main e-commerce images, Nano Banana 2.1 is also worth testing, particularly for the accuracy of product geometry, materials, and lighting.
For commercial integration, we recommend using your own product images as references and generating multiple images with each model at the same target resolution. Compare first-attempt success rates, final usability rates, and the actual cost per acceptable image before deciding which better suits your business.
We believe FLUX.3 Image is another model worth comparing with Flare, especially for image editing and brand asset production.
In a set of tests published on October 2, 2026, Fuser compared the two models using three text tasks and two local editing tasks.
For the text tests, FLUX.3 Image and GPT Image 2.5 Flare both correctly generated all 583 specified characters, with no noticeable difference observed.
However, FLUX.3 Image's performance in local editing tests deserves closer attention.
For example, when replacing a paper cup with a mug in a cafe photo, Fuser measured that 7.7% of pixels outside the specified editing region changed beyond the set tolerance with Flare, compared with only 0.4% for FLUX.3 Image. In another poster date revision, the proportions of changes in non-target regions were 25.0% and 12.9%, respectively. Fuser FLUX.3 Image hands-on comparison
We believe these figures show that, although Flare has improved its local editing capabilities, FLUX.3 Image may have an advantage in some scenarios that require strict preservation of the original image.
Our recommendation is this: if you mainly build general-purpose image generation, advertising creative, and rapid design tools, consider Flare first. If your application needs to make very small changes to finished posters, product photos, or brand images while keeping the remaining regions as unchanged as possible, prioritize FLUX.3 Image as a comparison model.
For the latter, what matters more to us is not whether the model generates attractive images, but whether it can accurately modify the specified region without disrupting previously approved content.
We recommend first signing in to RequestBeam and creating an API Key, then reading our documentation or asking AI to help you configure the connection.
We recommend completing a minimal generation test through the online tool first, then integrating according to RequestBeam's current GPT Image 2.5 Flare API documentation. This lets you verify whether the model meets your business needs before integration and helps avoid mistakenly carrying over endpoint paths and parameters from other providers into RequestBeam.
GPT Image 2.5 Flare's image quality parameter supports auto, low, medium, high, xhigh, and max, but this does not mean you can specify arbitrary dimensions.
According to OpenAI's official documentation, custom output dimensions must meet the following conditions:
In addition, generating images with more total pixels than 2560×1440 is currently an experimental capability. OpenAI image size configuration guidance
For an initial integration, we recommend using common dimensions such as 1024x1024, 1536x1024, or 1024x1536, and setting quality to medium or auto. Increase quality gradually after confirming successful calls.
Developers migrating older projects should also note that GPT Image 2.5 adds the xhigh and max quality settings, which earlier GPT Image models may not support. When using a third-party compatible endpoint, also confirm whether the platform fully supports these parameters.
When editing images with Flare, we recommend stating both what should change and what must remain unchanged in the Prompt.
For example, for product background replacement, do not just describe "change the background to a beach." Explicitly require preservation of the product's shape, label text, Logo, material, proportions, and camera angle.
OpenAI's official endpoint supports editing with multiple reference images and can also use a Mask to specify the region to modify. However, a Mask only provides a regional constraint for editing and does not guarantee that the model will follow the pixel boundaries exactly.
The official image editing endpoint allows up to 16 input images. Each input may be PNG, WebP, or JPEG and must be smaller than 50MB. OpenAI image editing endpoint documentation
To generate a transparent background, explicitly set background: "transparent" and choose PNG or WebP as the output format. JPEG does not support an Alpha transparency channel and is not suitable for storing transparent assets.
We recommend inspecting the actual final image file, especially problem-prone areas such as hair, glass, object shadows, and semi-transparent edges. If the image will enter a commercial design workflow directly, retain a human review or image post-processing step.
When returning an image, GPT Image 2.5 Flare's official Image API typically uses Base64-encoded content rather than directly providing a publicly accessible URL that remains available long term.
For example, in the official image generation endpoint's response, you can retrieve the image content from data[0].b64_json, then decode the Base64 into a PNG, JPEG, or WebP file. The response_format: "url" setting used by older endpoints does not apply to GPT Image models. OpenAI image generation endpoint reference
If your application needs to display images to end users, we recommend decoding them after your server receives them, saving them to your own object storage, and returning accessible image URLs to the frontend. This also makes it easier to manage history, image expiry, and CDN caching.
For generation latency, although Flare is faster than the older version, occasional long-running requests are still possible. OpenAI's image generation endpoint supports streaming partial images, allowing applications to show visual results before generation finishes, but partial images should not be treated as final completed images.
In a real integration, we recommend setting timeouts, failure retries, and concurrency controls according to the business scenario, and recording each request's actual duration and Token consumption. For timed-out requests with unknown results, do not retry indefinitely, as this may cause duplicate generation and additional charges.
If users have demanding response-speed requirements, we recommend first using lower quality for previews, then deciding whether to run high-quality generation based on the user's selection. Compared with using max for every request, this usually makes overall generation costs and waiting time easier to control. OpenAI image generation and streaming documentation
gpt-image-2.5-flarePOST /openai/v1/images/generationscurl -X POST "https://api.requestbeam.com/openai/v1/images/generations" \
-H "Authorization: Bearer <API_KEY>" \
-H "Content-Type: application/json" \
-d '{"model":"gpt-image-2.5-flare","prompt":"Create a product photograph of a nostalgic toy propeller airplane in clear blister packaging.\nGive the airplane rounded wings, a spinning nose propeller, and lightly worn painted-metal details.\nUse realistic plastic and metal textures, soft studio lighting, shallow depth of field, and crisp retail printing.\nInclude only the packaging label \"Christmas Memories Edition\". Make the design original, with no trademarks or watermarks.","n":1,"size":"1:1","quality":"low","output_format":"png","stream":false,"resolution":"1k","image_urls":[],"background":"auto"}'Family and protocol reference
GPT Image 2.5
Full API documentationGenerate images with Flare or edit reference images with Sunburst.
https://api.requestbeam.com/openai/v1/images/generationsUse gpt-image-2.5-flare for fast image generation or gpt-image-2.5-sunburst for precise editing. Check GET https://api.requestbeam.com/v1/models for availability with your API key.
POST https://api.requestbeam.com/openai/v1/images/generations
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json{
"model": "gpt-image-2.5-flare",
"prompt": "A red ceramic cup on a white table.",
"size": "1:1",
"resolution": "1k",
"quality": "low",
"n": 1
}| Parameter | Values |
|---|---|
model | Required: gpt-image-2.5-flare or gpt-image-2.5-sunburst |
prompt | Required text describing the image or edit |
n | Integer from 1 to 4; default 1 |
size | auto, aspect ratio, or pixel dimensions such as 1600x1200 |
resolution | 1k (default), 2k, 4k; ignored for exact pixel sizes |
quality | low, medium, high, xhigh, max, auto (default) |
image_urls | Up to 16 publicly accessible HTTP(S) image URLs |
output_format | png (default), jpeg, webp |
output_compression | Integer from 0 to 100 for JPEG or WebP |
background | auto, opaque, transparent; transparent requires PNG or WebP |
moderation | low (default), auto |
stream | false (default) for JSON, true for SSE completion events |
Supported aspect ratios: 1:1, 3:2, 2:3, 4:3, 3:4, 5:4, 4:5, 16:9, 9:16, 2:1, 1:2, 21:9, 9:21, 3:1, 1:3. Exact pixel dimensions must be multiples of 16, have no side above 3840 pixels, an aspect ratio at most 3:1, and 655,360 to 8,294,400 total pixels. For example, square 4k output is 2880x2880, not 4096x4096.
{
"model": "gpt-image-2.5-sunburst",
"prompt": "Keep the product and its lettering. Replace the background with a white studio.",
"image_urls": ["https://example.com/product.png"],
"resolution": "2k",
"quality": "xhigh"
}References use this JSON generation endpoint. Upload local images to your own storage first. Base64 data URIs, multipart uploads, masks, and a separate /images/edits endpoint are outside this model's public contract.
The HTTP connection waits for completion; clients do not poll an upstream task. JSON responses include final image URLs and available token usage:
{
"created": 1789035000,
"data": [{"url": "https://example.com/generated-image.png"}],
"usage": {"input_tokens": 15, "output_tokens": 196, "total_tokens": 211}
}With stream: true, wait for image_generation.completed and [DONE] on the same connection. Partial image previews are not guaranteed.
Billing uses actual text input, cached text input, image input, cached image input, and image output tokens. Both variants have the same token prices. The model catalog shows your current price and the separate OpenAI reference price. An image is not a fixed-price unit: prompt length, references, size, quality, and count affect the total. A temporary reservation is reconciled after completion. Choosing an explicit quality makes expected usage clearer.
Invalid parameters return 400; authentication failures return 401; insufficient balance returns 402. An unavailable model returns 404, rate limits return 429, and generation failures return 5xx. Do not automatically resubmit a request whose completion status is unknown.
See Error Codes and Troubleshooting for error codes, streaming failures, and retry boundaries.