OpenAIIMAGEGuest engine

GPT Image 2.5 on VisionX.

Two OpenAI siblings priced the same — Sunburst for fidelity, Flare for speed.

2models on rosterfrom 3.68 VXper imageup to 16reference imagesCast-readyidentity locked
01

What GPT Image 2.5 is on the roster

THE MODEL

GPT Image 2.5 is the current OpenAI image generation on the roster, and it arrives as a pair rather than a tier. Sunburst holds detail and reference likeness through a transformation; Flare answers faster. They cost the same per image, so the choice is about the shot in front of you, not the budget.

Both take a brief or up to sixteen reference images. The gain over the previous generation is in edits: change the one element you asked for and leave the subject, composition and lighting where they were — which is what makes it usable on a Cast frame instead of a fresh stranger.

Reference likeness survives the edit

Faces, distinctive features, lighting and texture stay recognisable after a transformation, so a Cast member still reads as the same person on the other side.

Edits stay local

Ask for one change and the surrounding subject, framing and treatment hold. Successive passes stack instead of quietly undoing each other.

Pick on latency, not on price

Sunburst for finals and precise retouching, Flare for volume and iteration — OpenAI positions Flare as the faster of the two. The composer quotes both before you run.

02

What GPT Image 2.5 actually does

CAPABILITIES

What the 2.5 pair changes, and how to pick between two models that cost the same.

Likeness survives the transformation

Subjects, distinctive features, lighting and texture stay recognisable after an edit. That is the difference between restyling your Cast member and generating a stranger who owns their jacket.

Reference fidelity

The edit lands where you pointed

Change the one element named in the brief and the surrounding subject, composition and visual treatment hold their position instead of drifting with it.

Local edits

Passes stack instead of fighting

Successive edits are less likely to undo earlier ones or soften the frame, so a shot can be walked to an approval note over several rounds rather than restarted.

Iteration

Sixteen references on one call

A Cast member, a product and a set plate can condition the same frame together — the deepest reference budget on the image roster, matched only by GPT Image 2.

Reference edits

Two speeds, one price

Sunburst holds the most detail and the most precise edits; OpenAI positions Flare as the faster of the pair for everyday volume. Neither is the budget option — the composer quotes them the same.

Sunburst vs Flare

Layout briefs still land

The instruction-following and legible on-image type that made GPT Image worth routing layout work to carry forward into the 2.5 pair.

Instruction following
03

Every variant, exactly as the studio runs it

SPEC SHEET

These rows are read live from the engine registry and the pricing engine. What you see here is what the composer quotes.

ModelModesReferencesOutputVX
GPT Image 2.5 SunburstGPT Image 2.5 SunburstImageup to 16 refsstills3.68 VX · per image, no references
GPT Image 2.5 FlareGPT Image 2.5 FlareImageup to 16 refsstills3.68 VX · per image, no references

Self-serve rates run as low as $0.085 per VX at scale. The composer quotes the exact VX for your shot, references included, before you run it.

04

How to write for GPT Image 2.5

PROMPTING

On an edit, name the change and name what must NOT move. 2.5 is markedly better at honouring the second half of that sentence than anything before it.

REFERENCE EDIT
Use [Image1] as the subject. Keep the face, pose, hair and
the blue backdrop exactly as they are.

Replace the red shirt with an ivory tuxedo jacket and black
bow tie. Match the existing key light and film grain.

Do not change the framing or the background gradient.

Write the preservation list

Spell out what stays — face, pose, lighting, background. The model is built to hold those, but only the ones you name are guaranteed attention.

One change per pass

Iteration is the strength here. Two passes each asking for one edit beat one pass asking for four, and 2.5 is far less likely to undo the earlier one.

Bind the Cast, then describe the change

Attach the reference and ask for the edit rather than re-describing the person. Identity carries through an edit far better than it survives a re-description.

Pick the sibling by deadline

Flare while you are exploring, Sunburst for the frame that goes to the client. Since the quote is the same, there is no reason to compromise on the final.

05

What teams route to GPT Image 2.5

USE CASES

Where the fidelity gain actually pays for itself.

Cast-accurate campaign frames

Put a Cast member in new wardrobe, a new set or a new season without losing the face the client signed off on.

Product retouching and variants

Swap a background, change a colourway or clean a hero shot while the product geometry and label copy stay exactly where they were.

Multi-round client revisions

Notes that arrive one at a time over a week — each pass applied on top of the last rather than a fresh render every round.

High-volume iteration

Flare for the exploration phase where you need forty options before lunch, then Sunburst on the two that survive.

Posters and packaging with real type

Headlines, ingredient panels and multilingual signage that come back spelled the way you wrote them.

Storyboard continuity

A sequence of frames that hold the same character, wardrobe and location from panel to panel.

06

GPT Image 2.5 against the alternatives

COMPARISON

Where the 2.5 pair sits against the rest of the image roster.

GPT Image 2.5 capability comparison against GPT Image 2 and Seedream 5.0 Pro
CapabilityGPT Image 2.5GPT Image 2Seedream 5.0 Pro
Best atReference-faithful editsType accuracy, literal layoutStylised campaign imagery
Reference images per callMost on the rosterMost on the rosterDeep
Repeated edit passesStrongestWorkableWorkable
Cost positionPremium — both siblings equalPremiumFlat rate per image

Qualitative routing guidance, not a benchmark. Exact reference caps and the VX quote for each engine are in the spec sheet above, read live from the registry and the pricing engine.

FAQ

GPT Image 2.5, answered

What is the difference between GPT Image 2.5 Sunburst and Flare?
Latency against detail, not price. Sunburst is the one to use for precise editing and the frame that ships; OpenAI positions Flare as the faster choice for everyday, high-volume generation. Both accept the same briefs and the same sixteen reference images, and VisionX quotes them at the same VX per image, so pick on the deadline rather than the budget.
How is GPT Image 2.5 different from GPT Image 2?
The gain is concentrated in reference work. Subjects, features, lighting and texture are more likely to stay recognisable through a transformation, edits are more likely to change only what you asked for, and repeated passes are less likely to undo each other or soften the image. GPT Image 2 stays on the roster, and so do 1.5 and 1.
Can it edit an image I already have?
Yes. Attach up to sixteen references — an upload, an asset from your library, or a Cast frame — and describe the change. Naming what must stay unchanged alongside what should change is what makes this generation worth routing edits to.
Which one should I put in a workflow?
Flare for any step that runs often or fans out into many variants, Sunburst for a final or approval-facing frame. Because the two are priced identically, moving a step from one to the other changes how long it takes and how tightly it holds a reference, not what it costs.

Try GPT Image 2.5 on your Cast.

20 VX free on signup, no card required.