GPT Image 2.5 Flare vs Sunburst: which model to use

GPT Image 2.5 is not one API model. OpenAI shipped two on September 8, 2026: Flare (gpt-image-2.5-flare) and Sunburst (gpt-image-2.5-sunburst). They share token rates, quality knobs, transparent-background support, and the same Image API / Responses API surface. They do not share a job description.

Flare is the default for most applications. Sunburst is the slower, tighter-control option for campaign stills and edit-heavy production. If you pick the wrong one, you either wait too long on drafts or spend a precision model on a thumbnail that only needed to be fast.

Neither model is in the OmniArt picker yet. GPT Image 2 is. Use this guide to decide which 2.5 model a brief is asking for, and which OmniArt model covers the same risk today.

Flare vs Sunburst at a glance

OpenAI's image-generation guide is blunt: choose Sunburst when editing precision matters most, and Flare for fast, high-quality everyday image generation. That is the whole product split. Everything else is a consequence.

What is actually the same

Do not invent a pricing or resolution gap that the docs do not contain.

Both models:

• Generate and edit from text and image inputs.
• Support the Image API (generations and edits) and the Responses API image-generation tool.
• Add xhigh and max quality on top of low / medium / high, and default to auto.
• Accept custom WIDTHxHEIGHT strings with the same constraints: multiples of 16, aspect between 1:3 and 3:1, edges at or below 3840px, 655,360 to 8,294,400 pixels, with sizes above 2560x1440 experimental.
• Support opaque and transparent backgrounds. Transparent requires png or webp.
• Stream partial images. Each partial costs an extra 100 image output tokens.
• Require organization verification before GPT Image use.
• Do not currently list Batch pricing, unlike gpt-image-2.

The pricing page lists identical standard rates for Flare, Sunburst, and GPT Image 2. OpenAI still warns that equal token rates do not mean equal cost per image, because token consumption can differ by model and quality setting. Read usage on real requests instead of assuming a 1K square at high costs the same on all three.

Where they diverge

The divergence is time versus control, not a different feature checklist.

Flare is a throughput model. OpenAI names creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation. Early-customer copy from Manus's evaluation team, published with the launch, said Flare produced high-quality images at two to four times the speed of GPT Image 2 in that team's tests, with better transparent-background generation. That is a customer evaluation, not OpenAI's 50% latency figure. Both are useful as hypotheses. Neither replaces a stopwatch on your prompt set.

Sunburst is an edit-stability model. OpenAI names production-ready campaign creative and polished product imagery. The API video calls it the more capable model for sharper detail, more natural lighting, richer textures, and closer instruction following. Higgsfield AI's launch quote is about preservation: character, composition, and visual identity surviving edits. That is the Sunburst job — the tenth revision should still be the same product, not a cousin of the first frame.

Adobe's launch quote is about distribution rather than the split: GPT Image 2.5 models are available in Adobe Firefly. That tells you the family is being wired into existing creative suites. It does not tell you which of the two models Firefly will call for a given tool.

How to choose for a real brief

Start from the failure you cannot afford, not from the model card.

A practical rule: draft on Flare, lock on Sunburst. Generate the composition and copy at a low or medium Flare setting. When the layout is accepted, rerun the same prompt and references on Sunburst at high or above. Change one variable — the model — so you can see whether the extra time bought anything you would actually ship.

If you are still on GPT Image 2, the same rule maps onto OmniArt: draft at 1K / Low, then spend 2K or High only on the accepted composition. See the GPT Image 2 prompt guide for that ladder.

API shape: Image API vs Responses API

The Flare/Sunburst choice is independent of the endpoint choice, but the endpoint changes the workflow.

• Image API — one prompt in, one image (or n images) out. Use it for a packshot, a transparent asset, or a single inpaint. Set model to gpt-image-2.5-flare or gpt-image-2.5-sunburst directly.
• Responses API — a mainline model calls the image tool inside a conversation. Use it when the next instruction is "now make the mug coral" rather than a full rewrite. You can force action: "generate" or action: "edit"; auto lets the tool decide. Forcing edit with no image in context errors.

For creators who never touch an API, ChatGPT Images 2.5 is the Responses-API idea with a canvas: Sketch, comments, and multi-turn chat. The what shipped article covers that surface. The prompt and editing guide covers how to write the turns.

What to use on OmniArt while 2.5 is rolling out

Match the risk, not the name.

The point of a two-model family is that most teams should not run one image model for every job. OmniArt already works that way: GPT Image 2 for control, Nano Banana 2 for photographic stills, Seedream 5.0 Pro for reference-heavy sets, then a video model once the still is approved.

FAQ

Is Sunburst always higher quality than Flare?

OpenAI positions it that way for detailed creative work, and it is the slower of the two. Quality settings (low through max) exist on both. A Flare max request is not the same object as a Sunburst low request. Compare at a fixed quality and size.

Do Flare and Sunburst cost different amounts?

Per-token rates are the same. Per-image cost can differ if one model emits more tokens at the same quality and size. OpenAI tells you to read usage rather than reuse the GPT Image 2 calculator blindly.

Can I use both in one workflow?

Yes. Draft on Flare, lock on Sunburst, and keep the prompt, references, size, and seed-equivalent constraints identical between the two runs. That is the cheapest way to learn whether Sunburst is earning its extra time.

Which one should a developer make the default?

OpenAI's default recommendation is Flare. Make Sunburst an explicit opt-in for jobs that already failed a preservation or typography check.

Getting started on OmniArt

You can rehearse the Flare/Sunburst split before the models land. Open GPT Image 2 on OmniArt and run the same brief twice: once at Low / 1K as the "Flare" stand-in, once at High as the "Sunburst" stand-in. Keep references and copy identical. If Low already passes, you do not need the expensive setting. If High is the first usable frame, that brief is a Sunburst candidate when 2.5 is available.

When the still is approved, take it into the video workspace from the same account. The image model that composed the frame and the video model that moves it do not have to be from the same lab.