GPT Image 2.5, in a fast lane and a fine one
Same model, two lanes, one price. Flare hands the picture back quickly. Sunburst takes longer and renders finer. Both write legible text into the image, hold a face across edits, read up to 16 reference images at once, and go up to 4K.
- Two lanes, 2 speeds, identical cost per resolution
- Up to 16 reference images in a single edit
- 1K, 2K and 4K output, with all 13 aspect ratios open at every tier
The choice
Pick a wait, not a budget
Most model tiers ask you to trade money for quality. GPT Image 2.5 does not. Flare and Sunburst cost exactly the same at every resolution, so the only thing you are choosing is how long you are willing to wait for the render. Switch lanes from the pill in the composer, mid project, without repricing anything.
Flare
The fast lane
Flare is the default. It returns the picture in the time it takes to read the prompt back, which is what you want when you are still deciding what the image should be. Draft a concept, throw it away, draft another. The whole point of a fast lane is that a bad idea costs you a minute instead of an afternoon.
Pick it when: you are exploring, iterating on a prompt, or generating in bulk and volume matters more than any single frame.
Sunburst
The quality lane
Sunburst spends longer on the same prompt and comes back with cleaner detail, steadier typography and fewer of the small structural mistakes that force a reroll. It is the lane for the frame that ships: the one going on the storefront, into the deck, or onto the printed poster.
Pick it when: the render is final, text has to be perfect, or you are going straight to 4K for print.
Both posters came from the same prompt at the same seed and the same resolution. The only difference is which lane rendered them.
Both lanes bill the same number of credits at 1K, at 2K and at 4K. Switching from Flare to Sunburst never changes what a generation costs you.
Two ways in
Start from a sentence, or start from pictures you already have. The model picks the right path from what you attach, so there is no mode to set first.
From a prompt
Describe the image and get it back. Prompts run to 20000 characters, which is enough to specify a full layout: the composition, the palette, and every word you want rendered into the frame, quoted so it comes back exact.
From up to 16 references
Attach images and the same call becomes an edit. Sixteen slots is enough to hold a product, a face, a logo and a style board in one request, so the output matches your brand instead of approximating it. Attaching anything switches the model to editing automatically.
One prompt, three sizes
Resolution is a pill in the composer, not a different model. The same prompt and the same references run at any of the three, and each tier is its own price so you only pay for the size you actually need.
1K
The working size. Fast, cheapest of the three, and the right default for social posts, thumbnails and anything you are still iterating on.
2K
The delivery size. Enough resolution for a full width hero, a slide, or a product page without paying for print detail you will never see on screen.
4K
The print size. For posters, packaging and billboards, or any time someone is going to crop hard into the frame and still expect it to hold up.
All 13 ratios, at every resolution
GPT Image 2 locks the two square-ish ratios out of its 2K and 4K tiers. GPT Image 2.5 does not: every ratio below is available at 1K, at 2K and at 4K, including the ultrawide and the tall ones. Pick the shape the placement needs rather than the shape the tier allows.
The default is auto, which lets the model choose a shape from the prompt. 4:5 is not part of this family; reach for Nano Banana 2 when a placement needs it.
Pricing
Credits, and nothing per seat
Every generation draws from one credit balance shared across every model on the platform. There is no per model subscription and no separate image plan. What a run costs depends on the resolution you picked, never on which lane you chose.
What people ship with it
GPT Image 2.5 is the one to reach for when the picture has words in it, or when the same face, product or logo has to survive a dozen variations.
Typography that reads
Headlines, price tags, packaging copy and slide text come back legible rather than as letter shaped noise. Quote the exact words in the prompt and they render as written.
Product photography
Attach the real product and get it lit, staged and placed in a scene it was never shot in, without the shape drifting between shots.
One face, many frames
Feed a portrait as a reference and it holds across an entire campaign, so a character stays recognisably the same person from the first frame to the last.
Packaging and mockups
Put a real label on a real box at 4K, close enough to send to a printer rather than to a reviewer.
Editing what exists
Change a background, swap a colourway, remove an object or extend a crop, working from your own file instead of regenerating from scratch.
Carousels and decks
Sixteen reference slots keep a multi slide set on one visual system, so slide nine still looks like it belongs with slide one.
The limits, up front
Worth knowing before you spend a run finding out.
20000 character prompts
Long enough for a full design brief with every visible word quoted.
16 references
Per request, across both lanes. A seventeenth image is rejected before anything is charged.
Up to 4K
Three tiers, each its own price. 1K is the default so an omitted setting never costs you more.
13 aspect ratios
Including auto. 4:5 is not among them, at any resolution or in either lane.
Questions
Time and finish. Flare returns faster; Sunburst spends longer and comes back with cleaner detail and steadier text. They are the same underlying model with the same inputs, the same limits and the same price at every resolution.
No. Both lanes bill identically at 1K, 2K and 4K. Switching lanes changes how long you wait, never what you pay.
Both are live and you can keep using either. GPT Image 2.5 is newer and adds the two lanes plus a wider set of aspect ratios that stay available at 2K and 4K. GPT Image 2 remains the cheaper of the two per generation.
Up to 16 in a single request, in either lane. Attaching any reference switches the call from generation to editing automatically, so there is no separate mode to select.
Yes. Every one of the 13 ratios is available at every resolution tier, squares included. That is a change from GPT Image 2, which refuses auto and 1:1 above 1K.
This family is one of the stronger ones for typography. Quote the exact words you want in the prompt and add a line saying they must read exactly as written; that phrasing alone removes most misspellings.
No, not in either lane and not at any resolution. Nano Banana 2 renders 4:5 and is the one to switch to when a placement requires it.
Yes. The two lanes are separate public model ids, gpt-image-2-5-flare and gpt-image-2-5-sunburst, and both are available over the REST API and the MCP server with the same aspect ratio and resolution parameters the studio uses.
Try both lanes on the same prompt
Run it through Flare, run it through Sunburst, and keep whichever one you would actually ship. They cost the same, so the comparison is free apart from the second render.