You stop re-editing the same frame. GPT Image 2.5 changes one thing. The rest stays exactly as it was.
Free account · ⚡50 on signup · No card
● The set
Four briefs, four jobs. Concept, cinema, narrative and a wide one.
GPT Image 2.5
“The peel opens. A full orchestra plays inside it.”
GPT Image 2.5
“A gallery, and one cube with its own weather.”
GPT Image 2.5
“Two people laugh under an umbrella. The city behind them does not.”
GPT Image 2.5
“Two swordsmen, wet stone, and a garden lit in neon.”
● What changed
01
It preserves the subjects in your reference photos.
02
It changes what you asked for. The rest of the frame holds.
03
Earlier edits stay consistent. Each new edit builds on the last.
04
It handles more complex layouts. Transparent backgrounds are among them.
05
It reflects a visual style more closely than 2.0 did.
06
More natural lighting. Richer textures out of the box.
● Two ways in
Write every clause. Subject, layout, style, words. OpenAI states it reflects a named visual style more closely. It handles more complex layouts. Transparent backgrounds are among them. Same rate card runs on Sunburst when the brief wants tighter control.
Attach the photo whose subject must survive. OpenAI states it preserves those subjects. It edits only what you asked for. Earlier edits hold across the turns that follow. A photoreal retouch belongs on Nano Banana instead.
● The sheet
Flare for most briefs. Sunburst when the edits have to hold tight. Below: one brief on each.
| Spec | GPT Image 2.5 Flare | GPT Image 2.5 Sunburst |
|---|---|---|
| OpenAI's own framing | The default choice for most applications | Built for premium visual workflows |
| Speed | Higher quality than GPT Image 2 at 50% lower latency | Tighter control across a run of edits |
| Cast it for | Creator and social content, product experiences | Production-ready campaign creative |
| And for | Visual search, rapid prototyping, high-volume generation | Polished product imagery |
| Where it runs | The OpenAI API | The OpenAI API |
Flare
GPT Image 2.5
“Cool side. Same nave, colder beams, more of the crowd.”
Sunburst
GPT Image 2.5
“Warm side. Candlelight holds and the lasers read red.”
Flare
GPT Image 2.5
“Same room, tighter stance, a different mass in frame.”
Sunburst
GPT Image 2.5
“Wide atrium, one mass centred over the pool.”
● Words in the frame
Brand names, headlines and label text. Set where the brief put them.
GPT Image 2.5
“Brand name set on the bottle. Rings and droplets around it.”
GPT Image 2.5
“Label copy in frame. Smoke, wet stone, one hard light.”
GPT Image 2.5
“Frozen water, black ground, the dial still readable.”
GPT Image 2.5
“Headline copy on the block. The pen draws the line itself.”
GPT Image 2.5
“Caddy, cup and leaf. Steam, and the light doing the work.”
GPT Image 2.5
“A product on a plinth at dusk, edge lit from under.”
GPT Image 2.5
“One shoe, planted. Glass, moss and leaf around it.”
GPT Image 2.5
“Eight pralines, eight worlds, one presentation box.”
● How to
01
Name the subject, the layout and the words. One clause each.
02
Reference photos ride along. It keeps the subject you gave it.
03
Change one thing per turn. The price shows before every render.
● Every format
Concept, interior, narrative and atmosphere. Landscape and portrait from one composer.
ConceptGPT Image 2.5
“Jungle, arctic, desert, ocean and moon, one cabinet.”
ActionGPT Image 2.5
“Brass gears, a hat, and someone already running.”
InteriorGPT Image 2.5
“A station concourse, and a skyline that does not belong to it.”
CraftGPT Image 2.5
“A jacket cut from weather, on the bench and on the stand.”
WideGPT Image 2.5
“Monks on the terrace. The ship holds over the peaks.”
InteriorGPT Image 2.5
“An ordinary sofa. Jupiter overhead, and Saturn beside it.”
WideGPT Image 2.5
“A red plain at dusk. Two silhouettes, one rover.”
SurrealGPT Image 2.5
“A concert instrument left in the ice, lit like a landscape.”
AtmosphereGPT Image 2.5
“A lamp still burning. A figure walking the seabed.”
NarrativeGPT Image 2.5
“Burgers on the grill. Tentacles on the skyline.”
AdventureGPT Image 2.5
“Two lanterns, one doorway, and a galaxy behind it.”
SurrealGPT Image 2.5
“Someone presses a dune. The crease is the point.”
CinematicGPT Image 2.5
“Candles, cut glass, and the guests nobody seated.”
CinematicGPT Image 2.5
“Cars under snow. One dancer, mid line.”
WideGPT Image 2.5
“Torches, rope bridges, and roots the size of a street.”
● The sheet
Vendor claims from the 8 September 2026 announcement. OpenClips defaults arrive with the catalog id.
● What it is cast for
The jobs OpenAI names. And the ones it wins.
● Quoted first
Per render · gpt-image-2.5-flare
43
$0.43 at pack rate
The default tier. Same price, sooner.
Per render · gpt-image-2.5-sunburst
43
$0.43 at pack rate
The quality tier, when the edits have to hold.
Priced per render, never per month. Both tiers cost the same, so the choice is speed against control. The composer shows the exact ⚡ on the button before you press it. A free account starts with ⚡50. Packs start at $10/1,000⚡.
● OpenClips MCP
No console and no syntax. Your agent briefs the model through the OpenClips MCP server. One sentence is the brief.
https://mcp.openclips.aiStep 01
One sentence is the brief. Subject, layout, style.
Step 02
The canvas and the ⚡ quote. Drafted and queued for you.
Step 03
Download it, remix it, or ship it. Credits settle automatically.
● Director's notes
Four habits that keep an edit an edit.
It edits only what you asked. Say that one thing.
One change per turn
Reference photos hold their subjects. Attach the one that matters.
Bring the reference
Each edit builds on the last. Quality holds over the run.
Keep the thread
It handles complex layouts. Transparent backgrounds are among them.
Say transparent
● From the blog
Studio NewsOne wallet, every model, and where the free signup grant goes.
Read it →
Prompt CraftWhich frame wins where, and how to compose for it before you render.
Read it →
Studio NewsHow the MCP door is built, and what an assistant may call through it.
Read it →● Questions
OpenAI's image model, announced on 8 September 2026 as its new state of the art. The pitch is sharper details, more precise editing and faster generation. In practice that reads as more natural lighting, richer textures, and reference photos whose subjects survive the render.
Yes. Both tiers ride the composer as of 9 September 2026. Flare is the default for most work. Sunburst is the tighter multi-turn tier. Same wallet, same library, same reference envelope as every other image model on the roster.
OpenAI states image generation latency is reduced by up to 50% compared with Images 2.0. That is the vendor figure, and it is the only speed number this page carries. Wall-clock time on OpenClips also depends on queue and canvas, and the composer reports the render as it runs.
That is the headline change. OpenAI states it is better at editing only what you asked for, keeping the rest of the details the same, even with more complex subjects and backgrounds. Say what changes, not what to preserve.
Yes. Across a longer conversation it follows specific editing instructions more reliably. Earlier changes are more likely to stay consistent, each new edit builds on the work already done, and image quality does not degrade over the run.
Yes. OpenAI states the model can handle more complex layouts including transparent backgrounds, and that it is better at reflecting visual styles. Useful when the frame has to sit on a product page or inside another layout.
Flare is the default choice for most applications, and OpenAI positions it for creator and social content, product experiences, visual search, rapid prototyping and high-volume generation. Sunburst is for premium visual workflows that want tighter control across edits, such as production-ready campaign creative or polished product imagery.
Any WIDTH×HEIGHT the size field accepts, within the vendor bounds: edges divisible by 16, longest edge up to 3840 pixels, aspect within 3:1, total pixels between 655,360 and 8,294,400. Seven presets sit in the composer for the common canvases: 1024×1024, 1024×1536, 1536×1024, 2048×1152, 2048×2048, 2160×3840 and 3840×2160.
One flat ⚡ per render, size and quality tier included. The composer quotes the exact number on the button before you press it. A free account starts with ⚡50.
OpenAI states it continues to use C2PA metadata and invisible watermarking. That is provenance data attached to the file, not a visible mark across the frame.
It stays live beside the 2.5 pair. Same composer, same wallet, same library. Pick it when you need the older model's exact behaviour on a run. The composer chip switches between the three without moving anything else.
● Version history
Newest first, straight off the roster.
image · Live on OpenClips
Sharper detail, more precise editing, faster renders.
image · Live on OpenClips
Precise instruction-based generation.
● The receipts
Every link goes to the vendor's own page.
Leave everything else exactly where it was.
Generate with GPT Image 2.5⚡50 on signup. No card. The price shows before every render.
Or roll with a sibling model
Most image models re-render the whole frame when you ask for one change. GPT Image 2.5 is better at editing only what you asked for, keeping the rest of the details the same, even with complex subjects and backgrounds.
That holds across a conversation. Earlier changes stay consistent, each new edit builds on the work already done, and quality does not degrade over the run. Your ⚡50 signup grant covers the first pass on a live model.
Sharper details, more precise editing, faster generation. More natural lighting and richer textures. Better preservation of the subjects in your reference photos. More complex layouts, transparent backgrounds included, and closer adherence to a visual style. Latency is down by up to 50% against Images 2.0.
Two models reach the API. Flare is the default for most work. Sunburst is the premium tier for workflows that need tighter control across edits.
Free-form size on the wire, with seven presets sitting on the field for the common canvases (1024×1024, 1024×1536, 1536×1024, 2048×1152, 2048×2048, 2160×3840, 3840×2160). Any WIDTH×HEIGHT works within the vendor's own bounds: edges divisible by 16, longest edge up to 3840, aspect within 3:1. One flat ⚡ per render, size and quality tier included.
Sketch, Templates and image comments are features of OpenAI's own app, not of OpenClips. What arrives here is the model.
Flare is the flagship version and the composer's default choice. Sunburst rides beneath it as a Variant. Same page, same rate card, same reference envelope, for briefs that want tighter control across a multi-turn edit. Every take lands in one library, and one balance covers the whole models roster.
● Where to go next
