GPT Image 2.5 on VisionX.
Two OpenAI siblings priced the same — Sunburst for fidelity, Flare for speed.
What GPT Image 2.5 is on the roster
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.
What GPT Image 2.5 actually does
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 fidelityThe 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 editsPasses 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.
IterationSixteen 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 editsTwo 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 FlareLayout 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 followingEvery variant, exactly as the studio runs it
These rows are read live from the engine registry and the pricing engine. What you see here is what the composer quotes.
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.
How to write for GPT Image 2.5
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.
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.
What teams route to GPT Image 2.5
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.
GPT Image 2.5 against the alternatives
Where the 2.5 pair sits against the rest of the image roster.
| Capability | GPT Image 2.5 | GPT Image 2 | Seedream 5.0 Pro |
|---|---|---|---|
| Best at | Reference-faithful edits | Type accuracy, literal layout | Stylised campaign imagery |
| Reference images per call | Most on the roster | Most on the roster | Deep |
| Repeated edit passes | Strongest | Workable | Workable |
| Cost position | Premium — both siblings equal | Premium | Flat 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.
Same Cast, same wallet, different look
Route a board line to GPT Image 2.5, then run the identical shot on another family and compare takes. Identity lives in your Cast, not in any one model, so switching engines never strands the campaign.