GPT Image 2.5 Flare vs Sunburst: Which API Model?
OpenAI did not ship a single gpt-image-2.5 API. It shipped two:
gpt-image-2.5-flare— fastest high-quality everyday generationgpt-image-2.5-sunburst— most capable model, built for editing precision
That split is the developer-facing version of ChatGPT Images 2.5. ChatGPT users do not pick Flare or Sunburst by name. API, Codex, and anyone calling the Image API or the Responses image generation tool do.
This comparison uses OpenAI’s model pages, image generation guide, prompting guide, and the September 8 launch post. It is not a PixVerse latency benchmark.
PixVerse now includes GPT Image 2.5 Flare and GPT Image 2.5 Sunburst as well as GPT Image 2. You can build the still in the PixVerse workspace or with the API, then run image-to-video. Upload an API result only when the still was not made in PixVerse. A 14-day launch offer made Flare and Sunburst available at no credit cost to eligible overseas PixVerse members. As of September 27, 2026, do not assume that offer is still active, and do not treat either model as free or unlimited. Check the credit price shown in your workspace. See PixVerse’s September 20, 2026 Flare vs Sunburst article and the GPT Image 2.5 model page.
Bottom Line
| If you need… | Start with |
|---|---|
| Drafts, social, search, high volume, A/B stills | Flare |
| Campaign finals, product pack shots, identity-preserving edits | Sunburst |
| A GPT Image 2 pipeline that already looks good | Flare, then keep Sunburst only if quality drops |
| A GPT Image 2 pipeline that still fails hard edits | Sunburst, then try Flare once quality is proven |
| ChatGPT Sketch / templates / comments | ChatGPT Images 2.5, not an API id |
OpenAI’s prompting guide is the cleanest decision rule: Flare is the small, speed-optimized model with quality comparable to GPT Image 2. Sunburst is the base, quality-optimized model with higher image quality than GPT Image 2. Both improve precise editing and subject preservation.
The marketing page is more aggressive on Flare. It says Flare delivers higher-quality images than GPT-Image-2 at 50% lower latency. When the two documents disagree in tone, use the prompting guide for production choices and treat the 50% figure as OpenAI’s stated maximum latency cut versus Images 2.0.
Spec Comparison
| Flare | Sunburst | |
|---|---|---|
| Model ID | gpt-image-2.5-flare |
gpt-image-2.5-sunburst |
| Snapshot | gpt-image-2.5-flare-2026-09-08 |
gpt-image-2.5-sunburst-2026-09-08 |
| OpenAI one-liner | Fast, high-quality everyday image generation | Most capable model for generation and editing |
| Developer guide role | Small / speed | Base / quality |
| Product-page use cases | Creator and social content, product experiences, visual search, rapid prototyping, high-volume generation | Production-ready campaign creative, polished product imagery |
| Quality settings | auto, low, medium, high, xhigh, max |
Same |
| Backgrounds | auto, opaque, transparent |
Same |
| Endpoints | Image generations and edits; Responses image tool | Same |
| Batch | Not listed at launch | Not listed at launch |
| Inpainting | Supported | Supported |
| Listed token rates | Identical to Sunburst and GPT Image 2 | Identical to Flare and GPT Image 2 |
Shared size rules: auto or WIDTHxHEIGHT, edges multiples of 16, max edge 3,840px, long-to-short ratio ≤ 3:1, total pixels 655,360–8,294,400. Recommended sizes in the docs include 1024×1024, 1536×1024, 1024×1536, 2048×2048, 2048×1152, 3840×2160, and 2160×3840. Outputs above 3,686,400 pixels (2560×1440) are experimental.
Same Price List, Not the Same Bill
Both models list:
- Image input $8 / cached $2 / output $30 per 1M tokens
- Text input $5 / cached $1.25 per 1M tokens
- Image-only output, so text output is not billed
OpenAI says those rates match GPT Image 2, and also says the GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption. The models can use different token counts at the same quality label. Community threads on launch day made the same point: list price per token is unchanged; tokens per image are undocumented relative to GPT Image 2.
Practical implications:
- Do not assume Flare is cheaper because it is faster.
- Do not assume Sunburst costs the same as GPT Image 2
highif you switch toxhighormax. - Measure accepted-image cost, including retries, on your own sizes.
Responses API calls also bill the mainline chat model (OpenAI’s examples use gpt-6-astra) on top of image tokens.
What Launch Partners Actually Said
OpenAI published four customer quotes. They are useful as workload hints, not as independent leaderboards.
- Higgsfield: Flare understands what not to change — character, composition, visual identity — which matters for film, UGC, and ads.
- Adobe: GPT Image 2.5 models are in Firefly; the emphasis is faster generation and resolution consistency through refinements.
- Manus: Flare was 2–4× as fast as GPT-Image-2 in their evaluations, with improved transparent-background generation for brand assets, decks, and websites.
- Runway: Low latency and quality fit how creators already work in Runway.
If your job looks like Higgsfield or Adobe campaign work, default to Sunburst for finals. If it looks like Manus or Runway iteration volume, default to Flare.
How to Call Each Model
Image API — set model on the generation or edit request:
const result = await openai.images.generate({
model: "gpt-image-2.5-flare", // or "gpt-image-2.5-sunburst"
prompt,
quality: "high",
size: "1024x1536",
background: "auto",
});Responses API — pick a mainline model at the top level, then set the image tool’s model:
const response = await openai.responses.create({
model: "gpt-6-astra",
input: "Generate a 16:9 first frame of a red kettle on a marble counter, no extra text.",
tools: [{ type: "image_generation", model: "gpt-image-2.5-sunburst" }],
});Lock a snapshot (gpt-image-2.5-flare-2026-09-08 or gpt-image-2.5-sunburst-2026-09-08) when you need behavior to stay still after launch week.
Keep quality explicit when you compare the two models. auto can pick different internal settings, which makes latency and cost comparisons noisy. OpenAI’s prompting guide also says the same quality label does not imply the same look or the same response time across Flare and Sunburst.
A Sensible Evaluation Order
- Pick 10–20 real jobs: a poster with exact copy, a product cutout, a face-preserving restyle, a local object removal, a transparent icon, a 16:9 video first frame.
- Run them on Sunburst at
mediumorhigh. - Run the same pack on Flare at the same quality and size.
- Score: instruction following, identity/product lock, text, unwanted changes, alpha edges, time to first accepted image, cost per accepted image.
- Raise quality (
xhigh,max) only on the jobs that still fail. - Put Flare on the paths that pass. Keep Sunburst on the paths that do not.
That is the same logic OpenAI recommends for migrating off GPT Image 2. It also maps cleanly onto a PixVerse pipeline: generate the still with the API model that passed, then animate in PixVerse instead of asking the image model to invent motion.
Where This Fits PixVerse
PixVerse is the video production surface. GPT Image 2.5 is OpenAI’s stills stack.
A typical split:
- Flare for exploring five thumbnail directions in a few seconds each
- Sunburst for the locked hero still (pack shot, character sheet, poster)
- PixVerse for turning that still into a 5–15s clip with V6, C1, Seedance, Kling, or another video model
On PixVerse, Flare and Sunburst can now make the still and send it into image-to-video without leaving the workspace. GPT Image 2 on PixVerse remains the path when a pipeline is pinned to GPT Image 2. Use ChatGPT when you specifically need Sketch or on-image comments; those tools are not API parameters.
FAQ
Can I use both models in one product?
Yes. Route drafts to Flare and finals to Sunburst. Keep the prompt identical so you can tell whether the model, not the brief, changed the result.
Is Sunburst always sharper?
OpenAI positions it as higher quality than GPT Image 2 and as the precision model. It does not promise that every prompt looks better, and xhigh/max are not free quality upgrades. Inspect the output.
Do Flare and Sunburst support transparent backgrounds?
Yes. Set background="transparent" and decode PNG or WebP while preserving the alpha channel. Check hair, glass, shadows, and edges. A painted checkerboard is not transparency.