GPT Image 2.5 Flare vs Sunburst on fal: Which to Pick, What It Costs, and What It Is For (October 2026)
OpenAI's GPT Image 2.5 arrived on fal on 8 September 2026 in two variants. Token-based pricing that works out to roughly five to twenty-five cents an image, up to 16 references, masks and transparent backgrounds, and why it suits covers more than face references.
OpenAI's GPT Image 2.5 reached fal on 8 September 2026 in two variants that carry code names rather than descriptions: Flare and Sunburst. Both are image-editing endpoints, and picking between them is harder than picking between most model pairs, because the names say nothing, fal's pricing page shows only a placeholder, and almost nothing about their differences is published. This comparison sets out what is documented, what is only reported, and the job each is best suited to.
What costs what
fal's pricing data shows a "$1 per unit" stub for both endpoints, which is a token placeholder rather than a price. The real cost follows OpenAI's token pricing, $30 per million output tokens, and varies with the quality level you request.
| Quality setting | Approximate cost per image |
|---|---|
| low | about $0.05 |
| medium | about $0.10 |
| high or max | up to about $0.25 |
Those figures are a range derived from the token rate and typical image sizes, not an invoice. For comparison: Nano Banana 2.1 is about $0.034 at 1K, Seedream 5.0 $0.027 to $0.0675, and FLUX.3 about $0.05 per 1K image. GPT Image sits at the expensive end, which is a reason to use it deliberately rather than for batches.
What both accept
As far as fal's listing reads:
- References: up to 16 input images.
- Mask: you can paint the region to change and leave the rest alone.
- Quality: low, medium, high or max.
- Background: transparent backgrounds are supported, which most editors do not offer.
Flare versus Sunburst
This is the honest part: we do not have a documented difference. The listing gives two names and a shared set of inputs. What we found:
- One third-party leaderboard, reporting on 9 September, ranks Sunburst first for text-to-image. That is a single source and a leaderboard, not a test of editing.
- Nothing we found says what Flare optimises for.
- We have not run either on our own material.
A reasonable default while that is unresolved: start with Sunburst on the strength of the leaderboard, and run the same prompt and references through Flare if the first result disappoints. Both are the same kind of cost, so a paired test of one prompt at medium quality is about 20 cents.
What it is for
Covers, thumbnails and posters. A thumbnail with a headline is where a premium image model's cost is easiest to justify, though we have not benchmarked text rendering across the models above.
Edits that must leave most of the image alone. The mask is the feature: replace a sign, swap a product label, remove an object, with the rest untouched.
Transparent assets. A cut-out you can place on another background.
Not for reference images that anchor a face in a video pipeline. Those are better served by editors we have used for it (Nano Banana, FLUX), and a premium price buys nothing there. We do not use GPT Image 2.5 for face references, and we have not tested how it handles them.
A practical pairing with video
The cover or the first frame of a short video benefits from a model that renders text well; the character reference benefits from one that holds a face. They are different models for good reasons. Make the cover with GPT Image, make the character with Nano Banana or FLUX, and animate the character with a video model (the Kling 3.0 Pro guide covers the usual choice).
Honest status
Everything here is fal's listing, OpenAI's token rate and one leaderboard, read between 9 September and 18 October 2026. The per-image ranges are our arithmetic. We will update this page if a paired test tells us how Flare and Sunburst actually differ.
A paired test that costs about forty cents
Write one prompt for a cover image with a short headline, supply the same two or three references to both variants at medium quality, and generate each once. Compare how accurately the text is spelled, how well the references are respected, and how the result looks at thumbnail size on a phone, which is where covers are actually seen. If one variant is clearly better for your style, use it for the series; if they tie, take the one the leaderboard prefers and move on. Do not spend more than a dollar deciding.
Common questions
What is the difference between Flare and Sunburst?
We found no documented difference. One leaderboard ranks Sunburst first for text-to-image; nothing we found says what Flare optimises for.
How much does an image cost?
It follows OpenAI's $30 per million output tokens, which works out around $0.05 at low quality and up to about $0.25 at high or max.
Does it support transparent backgrounds?
Yes, fal lists transparent backgrounds along with masks and up to 16 references.
Should I use it for face references?
No. We use other editors for that and have not tested how GPT Image 2.5 handles faces.
Why not use it for everything?
It is the dearest image editor in this group, so use it for covers and masked edits, not for batches.
How to do this without code: Everything between a photo and a finished video is handled by the tool, so you rarely need a separate image editor. In AutoHustle this is the Me in scene tool: a few photos of you and one sentence describing the scene (about 2 credits per second, quoted before you press Generate). See How to put yourself in any scene and Put yourself in a movie scene.
Related reading: Sora Is Gone: 5 Alternatives That Still Take Your Face (September 2026) · HappyHorse 1.1 Reference-to-Video: How Many Photos You Really Need and What It Costs (October 2026)
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| GPT Image 2.5 Flare edit | openai/gpt-image-2.5/flare/edit |
| GPT Image 2.5 Sunburst edit | openai/gpt-image-2.5/sunburst/edit |
fal shows a placeholder price; billing follows OpenAI's $30 per million output tokens. Up to 16 references, mask input, quality low to max, transparent background supported. Read on 18 October 2026.