The API ships two image models, and the names invite a wrong reading. gpt-image-2.5-flare is the fast default; gpt-image-2.5-sunburst is the editing-precision option and generates more slowly. Neither replaces the other.
The choice exists only in the API: nothing published describes a model switch inside ChatGPT. What follows is what was documented, what the two share, and four questions that settle it.
Two model IDs, one product
The launch post introduces both in one sentence: “For developers, we’re introducing two new models in the API.” There is no gpt-image-2.5. Both IDs carry the same generation of quality and editing behaviour, and OpenAI frames the split as speed against edit control, not better against worse. Product-level context: the GPT Image 2.5 overview.
What OpenAI documents about each
Both are quoted in full.
Flare: the fast default
“GPT-Image-2.5 Flare brings the same improvements in quality, editing, and speed to the API and is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency.”
The 50% figure is the only latency number OpenAI attaches to either model.
A second number carries Flare’s name but comes from a customer: Manus reports “two to four times the speed of GPT-Image-2” — a partner’s measurement, not a benchmark. Request detail: the GPT Image 2.5 API overview.
Sunburst: the precision option
“GPT-Image-2.5 Sunburst is built for premium visual workflows that benefit from tighter control across edits. Use it for creative and editing workflows like production-ready campaign creative or polished product imagery.”
The same paragraph gives Sunburst its one documented cost: longer generation times, no number attached. “An extra level of precision” is undefined too, with no accuracy figure and no list of edits it holds better.
Where the two models are identical
Most of what you need is identical; this is the whole published comparison.
| Item | Flare | Sunburst |
|---|---|---|
| Published model ID | gpt-image-2.5-flare | gpt-image-2.5-sunburst |
| What OpenAI says it is for | Default for most applications | Premium workflows needing tighter edit control |
| Generation time | 50% lower latency than GPT-Image-2, official | Longer generation times, no figure published |
| Documented use cases | Social and creator content, product experiences, visual search, high volume | Campaign creative ready to ship, polished product imagery |
| Where you select it | The model field of an API request | The model field of an API request |
| Available in ChatGPT | No per-model selection documented | No per-model selection documented |
Both IDs are API-only and chosen the same way, and neither is something a reader can pick in the app. How much slower Sunburst is, or how much better its edits hold, is not stated.
How to choose between them
Four questions, ordered so the first removes the most uncertainty.
- Is the bottleneck throughput or edit stability across turns? Nothing gets re-edited: the fast lane. Three or four changes before shipping: the precision lane.
- Can you name the single detail the result must keep? Tighter control across edits pays off only when something specific has to hold, such as a product silhouette.
- One delivery or a batch? A batch multiplies every second of generation time by the image count; a single hero image does not.
- Will the edits happen in one session? Following instructions across turns is this generation’s documented strength.
Two worked cases: batch and refine
One case per lane, both runnable as written.
Case one: a batch of social frames
Batch work is Flare’s territory, and only one variable moves between frames.
Task: four social frames from one input, one variable at a time.
Input: a matte ceramic coffee mug on a plain warm-grey table.
Fixed in all four: 1:1 square, eye level, soft daylight from the left, generous margin.
Variable: frame 1 blue backdrop, frame 2 green, frame 3 terracotta, frame 4 charcoal.
Hold constant across frames: mug shape, handle position, shadow direction.
Exclusions: no text, no logo, no hands, no props, no collage, no watermark.
Output: four separate square images.Change only the backdrop word and run it four times. If the mug’s shape or the shadow direction moved, a constant you declared did not survive the batch.
Case two: refining one product photo
The path Sunburst is documented for: produce a version, change one condition, leave the rest.
Step 1 — one product photo: a leather notebook on a pale oak desk, soft daylight from the left,
45-degree three-quarter view, plain background, no text, no hands, no props.
Step 2 — change one condition only: replace the pale oak desk with brushed steel. Keep the
notebook, the light direction, the camera angle and the background tone unchanged.
Check after step 2: notebook colour, light direction, camera angle, background tone.To compare the models reproducibly, send the same prompt twice with every other field identical — same size, same quality and format values — and change only model, to gpt-image-2.5-flare and then to gpt-image-2.5-sunburst. Read the third condition, not the first: if it survived in one run but not the other, the extra time bought you something.
What OpenAI has not published
Four gaps matter; anything beyond them did not come from the vendor.
- A price for either model. No per-image cost was published for either; the launch post points at the pricing page.
- Any latency number for Sunburst. “Longer generation times” is words, not a percentage or a baseline.
- Quality tier names. Neither model has named grades, so “higher quality” and “extra precision” cannot be ranked against each other.
- A quantified editing comparison. Both are described as improving edits; neither is measured against the other.
Those four bound what can be checked; anything else is a third party’s number.
Sunburst vs Flare FAQ
- Which one should I call? Start with
gpt-image-2.5-flare, the documented default for most applications. Move togpt-image-2.5-sunburstwhen one specific edit must hold across several turns. - Can I choose the model inside ChatGPT? No documented switch exists; the help documentation never mentions the two IDs, so this stays an API decision.
- Can a prompt decide which model answers? No. The model is set in the request, not in the prompt text.
- How many comparable performance numbers are public? Two, and neither is a cross-model benchmark: Flare’s official 50% latency reduction against GPT-Image-2, and one customer’s two-to-four-times statement.
Where to go next
Three pages continue from here. The GPT Image 2.5 changelog separates model changes from app changes. The free access and limits guide covers what the plans include. The GPT Image 2.5 alternatives page sets both IDs beside the other generators.