Connect to GPT Image 2.5 Sunburst 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-photorealism-gpt-image-2-5-sunburst.webp"
}
]
} Sign in to run with your account balance.
Example preview
{
"data": [
{
"url": "/examples/gpt-image-25-extract-product-gpt-image-2-5-sunburst.webp"
}
]
} Extract the product from the input image and isolate it on a fully transparent background. Output: centered product, crisp silhouette, no halos/fringing. Preserve product geometry and label legibility exactly. Add only light polishing. Do not add a solid backdrop, checkerboard, scenery, or shadow. Do not restyle the product; remove the background and preserve clean alpha transparency.
Turn this drawing into a photorealistic image. Preserve the exact layout, proportions, and perspective. Choose realistic materials and lighting consistent with the sketch intent. Do not add new elements or text.
In this room photo, replace ONLY the white chairs with chairs made of wood. Preserve camera angle, room lighting, floor shadows, and surrounding objects. Keep all other aspects of the image unchanged. Photorealistic contact shadows and fabric texture.
GPT Image 2.5 Sunburst is an image generation and editing model released by OpenAI on September 8, 2026. It is also the version in the GPT Image 2.5 family that places greater emphasis on image quality and editing precision.
We believe the biggest change in Sunburst is not simply its ability to generate more polished images, but its ability to understand more accurately what users want to change and what should remain unchanged.
For example, when creating a product advertisement, you may only want to change the product's angle, adjust the background lighting, or replace a piece of text, without changing the product's appearance or the composition of the entire image. Sunburst is optimized for this kind of work that requires precise control.
OpenAI positions Sunburst as a model for advanced creative work and detailed image editing. Compared with Flare in the same family, it is more willing to spend time generating an image to improve output precision.
We recommend testing Sunburst first if your business primarily depends on high-quality final images, such as commercial photography, advertising design, and brand assets. But if you simply want to generate large numbers of images quickly, Flare may be more suitable.
GPT Image 2.5 Sunburst is not just a Text-to-Image model. It can also take existing images as input and use text instructions to edit images, transfer styles, and combine multiple reference images.
We believe this capability is closer to real design work than text-to-image generation alone, because users are usually not satisfied with the very first image and need to keep refining details.
For example, you can upload a character, an indoor background, and a clothing reference image, then ask the model to regenerate the scene with the specified positions and relationships.
In tests conducted by Genflick in September 2026, the GPT Image 2.5 family successfully completed a character-and-scene composition task involving three reference images, preserving the character's clothing, background features, and specified left-right positions.
Of course, this does not mean complex compositions will never go wrong. We recommend providing clear original images and explicitly stating which features the model should preserve when product appearance, character identity, or brand logos are involved.
Genflick multi-reference-image test
GPT Image 2.5 Sunburst supports six quality settings: low, medium, high, xhigh, max, and auto, with auto as the default.
For resolution, the official recommendations are 1024x1024, 1536x1024, and 1024x1536, with custom resolutions also supported when they meet the dimension constraints. The model supports outputs of up to approximately 4K-scale pixel counts, although higher resolutions still have experimental limitations.
We believe the greatest value of quality settings is that they let developers control costs according to business needs, rather than using max for every scenario. As we have mentioned repeatedly in our discussions of other models, the highest reasoning setting may not produce the best result.
For example, medium or high may be sufficient for ordinary article illustrations on a website. For commercial images that need to show product details, you can further test xhigh and max.
Note that a higher quality setting does not mean every requirement will be handled better. Some practical tests have even found that higher settings add decorative elements the user did not request.
We recommend establishing the composition and content at a lower-cost setting, then increasing the quality for the final assets. This is a more reasonable approach.
OpenAI image generation parameter documentation
Through RequestBeam, we want to help you experience GPT Image 2.5 Sunburst in real business scenarios sooner, rather than first spending considerable time integrating an API.
Simply sign in to RequestBeam, create an API Key, and add a small amount of credit to start evaluating the model with the online tools. You do not need to write API code first or buy a fixed plan for a short-term test, and you can reduce the hassle of managing accounts across multiple model platforms yourself.
We believe precise editing is GPT Image 2.5 Sunburst's most noteworthy capability.
A relatively common problem with previous image generation models was that a user wanted to change one small detail, but the model also changed a person's face, the product's shape, or even the entire background.
Sunburst is optimized for this situation. OpenAI specifically emphasizes its ability to perform local edits within complex subjects and backgrounds while preserving the original subject, composition, and brand elements as much as possible.
One of our e-commerce customers shared a comparison of Sunburst and Flare. Using the same product prompt across multiple quality settings, he found that Sunburst more consistently preserved the product's perspective, reflections, and composition, while Flare was more likely to rework the image under different settings.
We believe this capability is well suited to e-commerce product images, advertising assets, and local modifications to existing designs.
Character consistency has always been an important issue for users who create comics, story illustrations, and marketing assets with AI.
We are more interested in whether Sunburst can preserve a character's key features, such as facial structure, hairstyle, accessories, and overall visual style, while changing their pose, clothing, or background.
At launch, OpenAI specifically demonstrated its ability to move a character from a reference image into different scenes, adjust their appearance, and make successive edits.
One of our short-form drama customers told us that prompts that previously failed to produce satisfactory results improved noticeably with the Sunburst API, particularly praising character consistency and image quality.
We believe Sunburst is suitable for producing multiple promotional images of the same character, sequential story illustrations, and brand character assets that need to maintain a consistent visual identity.
Our recommendation is to first establish an approved character reference image, then use it to generate different scenes. Ensuring consistency in the base character assets before generating the finished images will lead to better results.
We believe commercial product photography is another important application for Sunburst.
In GPT Image 2.5, OpenAI focused on improving natural lighting, material textures, and the preservation of reference-image details. Sunburst puts further emphasis on high-quality final visuals.
For example, one of our e-commerce customers shared that he was generating promotional images for a skincare product. He uploaded a reference image to Sunburst to generate product visuals with different backgrounds while requiring the bottle proportions, packaging colors, and product logo to remain unchanged. He was very happy with the results.
Of course, evaluating commercial photography is not just about whether an image looks attractive. Everypixel conducted a blind test across 48 commercial photography scenarios in September 2026. Sunburst's photorealism preference share against Qwen Image 2 was 69.0%. Against GPT Image 2, however, it received only 52.7%, with a confidence interval that included 50%, so this cannot be considered a clear advantage in photorealism.
We believe this result is a useful reference. Sunburst's strengths may lie more in following instructions and controlling edits, rather than making every image look noticeably more realistic.
For e-commerce product images, we recommend checking product shape, packaging text, plausible lighting and shadows, and background details separately. E-commerce depends on the actual product matching its presentation. A mismatch can damage users' trust in your brand.
Everypixel commercial photography blind test
In advertising posters and marketing images, the model needs not only to generate an image, but also to handle text, numbers, and layout correctly.
We believe Sunburst already has considerable practical value in this area.
Feedback from several customers has told us that Sunburst delivers satisfying results in specific business tasks such as posters and nine-panel storyboards, significantly reducing the number of revisions and sometimes getting it right on the first attempt.
Based on feedback from these specific business uses, we believe Sunburst is already worth using in real work for promotional images with a small amount of clearly specified text. However, images with extensive text, complex layouts, or strict brand requirements still need human review.
Our recommendation is to clearly specify the text, its placement, relative sizes, and elements that must not be added. A structured Prompt is preferable, as this can help you generate better images.
We care more about whether Sunburst can fit into an actual design workflow than whether it can simply generate one good-looking image.
In commercial work, designers usually need to generate a first draft, then repeatedly revise it based on feedback, such as adjusting product placement, changing background colors, replacing text, and producing different versions for different channels.
The GPT Image 2.5 family improves detail preservation during successive edits, while Sunburst places greater emphasis on editing precision.
One of our customers, a developer who has used the GPT Image API for a long time, said that he had previously generated over approximately 80,000 images and observed improvements in reference-image use and interface elements compared with GPT Image 2.5 when testing the new version.
We believe the real question for marketing teams is whether fewer revisions are needed to get from the first draft to the final asset.
In this section, we refer to test data published by Artificial Analysis, Arena, Everypixel, and Quantslant, along with some experiences shared by customers.
On Artificial Analysis's AA-Image-T2I v2.0 leaderboard, GPT Image 2.5 Sunburst scored 1198 Elo at the max setting, placing first at the time.
On the same leaderboard, GPT Image 2.5 Flare Max scored 1191 Elo, while GPT Image 2 High scored 1172 Elo.
This evaluation uses anonymous pairwise image comparisons and calculates Elo scores from human reviewers' preferences, making it closer to users' judgments of overall generated-image quality.
However, we note that Sunburst's and Flare's 95% confidence intervals overlap, which means their overall text-to-image quality is relatively close.
We recommend testing Flare first if your business mainly involves ordinary text-to-image generation, then including Sunburst in the comparison for assets with higher final visual quality requirements.
Artificial Analysis text-to-image leaderboard
Elo · Higher is better
Sunburst and Flare have overlapping 95% confidence intervals; the scores do not establish a clear quality gap.
Source: Artificial Analysis · Zero-based scale
On Artificial Analysis's AA-Image-Editing v2.0, Sunburst Max scored 1183 Elo, compared with 1163 Elo for Flare Max and 1125 Elo for GPT Image 2 High.
We believe this result reflects Sunburst's product positioning more clearly than its text-to-image score.
Image editing is not about generating an image from scratch, but about producing a result that meets the user's expectations based on an existing image and requested changes. According to data shared with us by customers, the ratio of editing to generating from scratch is approximately 3:1. This set of results therefore points to a relatively clear advantage for Sunburst in image editing.
Our recommendation is to prioritize testing Sunburst if you are developing an AI image editor, a product background replacement tool, or an inpainting tool, rather than choosing a model solely on its text-to-image ranking.
Artificial Analysis image editing leaderboard
Elo · Higher is better
Preference scores measure the evaluated editing tasks, not a guarantee that every requested detail is preserved.
Source: Artificial Analysis · Zero-based scale
On Arena's Multi Image Edit leaderboard, GPT Image 2.5 Sunburst scored 1535±9, Flare scored 1501±8, and GPT Image 2 Medium scored 1454±4. These results were still marked as preliminary at the time.
We believe this score is particularly important for applications that need to combine multiple reference images.
For example, when combining people, clothing, and scenes from different images into a single image, the model needs not only to understand each image but also to handle the positions, proportions, and visual relationships among multiple objects.
However, Arena scores come from user preference comparisons and do not mean every object has been reproduced accurately. For product combinations, group portraits, and complex scene generation, we therefore still recommend treating human review of details as an indispensable step.
Arena multi-image editing leaderboard
Points · Higher is better
These results were marked preliminary. User preference scores do not mean every object was reproduced accurately.
Source: Arena · Zero-based scale
Everypixel tested 48 commercial photography scenarios, with anonymous reviews by 13 production team members.
In the comparison between Sunburst and GPT Image 2, the results across three main dimensions were: a photorealism preference share of 52.7%, prompt adherence of 55.2%, and aesthetic preference of 56.7%.
We believe this test reminds us that the 52.7% result does not establish a statistically clear advantage in photorealism. Although Sunburst has a slight edge in prompt adherence and aesthetics, those advantages are also relatively limited.
We therefore recommend that teams already using GPT Image 2 for commercial photography assets should not rush to migrate everything. Start by testing Sunburst on scenarios that require precise editing, complex compositions, and text handling, then decide whether to switch.
Everypixel 48-scenario blind test
Preference share (%) · Higher is better · 50% = equal preference
The realism confidence interval includes 50%: no statistically clear realism advantage. Other reported preference gains are modest.
Source: Everypixel · Zero-based scale
Quantslant tested Sunburst and Flare for generation speed and cost using the same prompt.
At a resolution of 1536×864, the results were as follows:
| Quality | Sunburst time | Flare time | Cost per image |
|---|---|---|---|
| Medium | 14.0 seconds | 12.8 seconds | $0.0086 |
| High | 27.8 seconds | 17.9 seconds | $0.033 |
| Max | 77.8 seconds | 37.6 seconds | $0.130 |
These data come from a specific test environment, with only a small number of samples at Medium. They should not be treated as stable platform response times or fixed pricing.
We believe the most important point here is that Sunburst and Flare use the same official Token rates, but Sunburst takes noticeably longer at higher quality settings. Moving from High to Max also increases the image cost considerably, without necessarily delivering a proportional improvement in image quality. We have emphasized this repeatedly: higher reasoning settings take more time but do not necessarily produce the best results. This is a counterintuitive point we keep stressing to our customers.
We recommend starting with High when evaluating ordinary images. Try Max for high-value advertising assets or final images that need to preserve complex details. For users who need quick generation and repeated previews, we recommend Flare more strongly.
Quantslant original speed and cost test
Seconds · Lower is better
Specific test conditions; Medium had few samples. These are not guaranteed platform response times.
Source: Quantslant · Zero-based scale
USD per image · Lower is better
Costs from this specific test, not fixed pricing or RequestBeam prices. Higher cost does not imply a proportional quality gain.
Source: Quantslant · Zero-based scale
We believe Sunburst and Flare are the most natural models to compare because they belong to the same product generation, but they are actually optimized for different business needs.
According to independent testing by Artificial Analysis, the difference in their text-to-image quality is not large. Sunburst Max scores 1198 Elo, while Flare Max scores 1191 Elo.
In image editing, however, Sunburst's advantage is more pronounced, with scores of 1183 and 1163 Elo respectively.
For speed, Quantslant's testing found that Flare was noticeably faster at High, with the gap widening further at Max.
We do not see these two models as a simple upgrade-and-replacement relationship, and their use cases are not identical.
If you are developing a blog illustration generator or a general AI text-to-image website, or need to generate social media assets in bulk, we recommend prioritizing Flare.
If your product requires users to upload reference images and repeatedly make precise edits to people, products, and local areas of an image, we recommend Sunburst more strongly.
For platforms with both generation and editing features, we recommend using Flare for quick creation and reserving Sunburst for detailed refinement. For both Flare and Sunburst, we recommend starting tests at high.
Artificial Analysis text-to-image comparison · Editing comparison
GPT Image 2 is an important reference point for Sunburst, because many applications already using OpenAI's image generation capabilities need to decide whether upgrading is worthwhile.
In Artificial Analysis's evaluations, Sunburst has higher Elo scores than GPT Image 2 in both text-to-image generation and image editing.
However, Everypixel's commercial photography blind test provides another perspective. Sunburst did not demonstrate a statistically clear lead in photorealism.
We believe this suggests the most noteworthy part of the new version is its handling of complex requirements and editing tasks, rather than an ability to generate noticeably more attractive images in every commercial scenario.
The quality settings also differ between the two versions. GPT Image 2 supports up to High, while Sunburst adds XHigh and Max, providing more configuration options for final outputs.
For teams already using GPT Image 2, we recommend prioritizing the migration of image editing and multi-reference-image composition tasks. For product showcase images and article illustrations that already perform consistently, you can keep the old model for comparison first.
We do not recommend replacing every generation task simply because Sunburst is newer. Deciding whether to migrate after testing your business use cases is the right approach.
Nano Banana 2.1 is Google's image generation and editing model and a cross-platform option worth considering when evaluating Sunburst.
On Artificial Analysis's AA-Image-T2I v2.0, Sunburst Max scores 1198 Elo, while Nano Banana 2.1 scores 1160 Elo.
On the image editing leaderboard, they score 1183 and 1137 Elo respectively.
These two results give Sunburst the advantage in independent preference tests.
However, cost is another factor to consider. Under the corresponding standardized evaluation settings, Artificial Analysis lists Sunburst at approximately $210.7 per thousand images and Nano Banana 2.1 at approximately $33.6 per thousand images. These figures apply only to those evaluation configurations, not as fixed costs for arbitrary resolutions and quality settings.
We believe you can prioritize testing Sunburst if you want higher image editing quality. If your business needs to generate large numbers of images, Nano Banana 2.1's cost advantage deserves close evaluation.
Artificial Analysis text-to-image and editing leaderboards
MAI-Image-2.6 is Microsoft AI's image generation model and is also worth including in model selection tests for commercial image generation.
On Artificial Analysis's text-to-image leaderboard, Sunburst Max scores 1198 Elo, compared with 1151 Elo for MAI-Image-2.6. On the image editing leaderboard, their scores are 1183 and 1139 Elo respectively.
Both tests show Sunburst leading in human preference scores. For cost, the standardized generation costs per thousand images in the same evaluation are approximately $210.7 and $38.9 respectively.
We believe Sunburst deserves priority consideration if your business needs premium advertisements, detailed product images, or brand assets that undergo multiple rounds of editing.
But if you need to generate background images in bulk for a large product catalog, or continuously produce large numbers of illustrations for a content website, MAI-Image-2.6 may be more economical.
Our recommendation for applications with large output volumes is to first define the minimum acceptable image quality, then choose the model that meets that standard at a lower total cost, rather than simply using the top-ranked model.
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 Sunburst 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.
Sunburst is designed to place greater emphasis on precision, so generation can take noticeably longer than Flare at high quality settings.
OpenAI officially supports six quality settings: low, medium, high, xhigh, and max represent different levels of investment in quality, while auto lets the model choose automatically.
We recommend against making max the default for every request.
For general business use, start with medium or high. Try higher quality settings if the output needs to be displayed at a larger size, preserve complex product details, or serve as high-value promotional material.
For cost, official standard pricing is $5 per million text input Tokens, $8 per million image input Tokens, and $30 per million image output Tokens. The actual cost per image also depends on resolution, quality, and Token usage.
After API integration, we recommend recording quality, size, response time, and actual usage together, then using your own business data to identify the most cost-effective parameter combination.
OpenAI official pricing and quality documentation
In OpenAI's native Images API, GPT Image 2.5 Sunburst returns Base64-encoded image data, usually in data[0].b64_json in the response.
This is not a generation failure. It is the data format returned by the official API.
We recommend decoding the image content after receiving the response, saving it to your own object storage, and then providing the frontend with an accessible image URL.
If you integrate through RequestBeam, you also need to check the response structure for the protocol you actually select, because a compatible service may wrap the image results.
The model also supports output formats including PNG, JPEG, and WebP. For a transparent background, set background to transparent and use PNG or WebP.
For ordinary website images, we recommend testing WebP or JPEG to reduce transfer and storage overhead. For transparent product assets, PNG or WebP with an alpha channel is more suitable.
OpenAI image output format documentation
If users need to keep making changes to the same image, we recommend first exploring the image generation tool in OpenAI's Responses API.
The official Images API is suitable for one-off generation or editing, while the Responses API is better suited to retaining context during a conversation and using image inputs to continue with subsequent changes.
When using the Responses API, you need to send the request through a main model that supports the image generation tool, then configure the image_generation tool to use gpt-image-2.5-sunburst, rather than using Sunburst directly as an ordinary text model.
For products that need reliable editing, we recommend retaining every generated image, original prompt, and editing instruction so users can recover earlier versions after multiple rounds of changes.
Also note that multi-turn editing does not guarantee that all content will remain completely consistent. For key elements such as faces, product packaging, and brand logos, we still recommend the necessary quality checks after every round.
gpt-image-2.5-sunburstPOST /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-sunburst","prompt":"Create a photorealistic candid photograph of an elderly sailor standing on a small fishing boat.\nHe has weathered skin with visible wrinkles, pores, and sun texture, and a few faded traditional sailor tattoos on his arms.\nHe is calmly adjusting a net while his dog sits nearby on the deck. Shot like a 35mm film photograph, medium close-up at eye level, using a 50mm lens.\nSoft coastal daylight, shallow depth of field, subtle film grain, natural color balance.\nThe image should feel honest and unposed, with real skin texture, worn materials, and everyday detail. No glamorization, no heavy retouching.","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.