AIREITERGPT-Image 2.5 vs GPT-Image 2: faster stills, edits that stop drifting
GPT-Image 2 already followed prompts and drew type well. The complaint after that was wait time, and faces or products that slid around after the second or third edit. OpenAI shipped 2.5 on 8 September 2026 for those two gaps — and split the API into Flare and Sunburst instead of one successor ID.
The wait is the obvious upgrade
OpenAI says image latency is up to 50% lower than Images 2.0, and that Flare is higher quality than GPT-Image 2 at that shorter wait. Manus told OpenAI their own evals saw Flare two to four times faster than GPT-Image 2. If your pipeline currently sits on GPT-Image 2 and the stills are already good enough, Flare is the first thing to try — not a new prompt style, just less idle time between candidates.
Edits stop rewriting the whole frame
Images 2.5 is built to change only what you asked for. OpenAI's demo is a living room that stays the same room while the furniture and light change; another is swapping a jacket without moving the person. Higgsfield described the same behavior: character, composition, and visual identity survive the edit. That is the gap GPT-Image 2 left on production jobs.
Lighting, texture, and infographics
The launch post calls out more natural lighting and richer textures, and better preservation of people in reference photos. The system card is more specific on graphics: infographic accuracy and layout improved versus Images 2.0. If GPT-Image 2 already draws a headline but breaks the grid, 2.5 is the model OpenAI is pointing at.
What actually changed from GPT-Image 2
Same family, same token price list, not the same model. OpenAI still lists GPT-Image 2 for older integrations and tells new work to start on Flare or Sunburst.
One model became two jobs
GPT-Image 2 is a single ID. 2.5 is Flare (default, speed) and Sunburst (precision, slower). There is no bare gpt-image-2.5. LMArena's image arenas put Sunburst and Flare at #1 and #2 after launch, with Sunburst up 81 points on multi-image edit and 40 on text-to-image versus GPT-Image 2 medium — that is Arena's board, not a test we ran here.
Quality ladder went past "high"
GPT-Image 2 tops out at high (plus auto). 2.5 adds xhigh and max. Official per-million rates match GPT-Image 2. The image generation guide now estimates 2.5 output tokens by model, quality, and size — and says the same quality label can spend a different token count than GPT-Image 2.
Keep 2 only if the brief already ships
OpenAI's own migration note: if GPT-Image 2 already meets quality, try Flare for latency; if 2 smears a face, logo, or product edge across edits, try Sunburst. GPT-Image 2 stays live at /image/gpt-image-2. Do not rewrite working prompts until an A/B with the same references actually wins.
Flare or Sunburst after you leave GPT-Image 2
Flare — throw away more candidates per hour
Creator and social crops, product-page drafts, visual search, anything you will generate fifty times before you keep one. Adobe put the 2.5 models into Firefly the day they launched; that is the "default app model" slot OpenAI is describing. On this site the ID is gpt_image_2_5_flare.
Sunburst — one change, rest of the pixels locked
Brand stills, retouched pack shots, a label rewrite on a bottle that must stay the same bottle. OpenAI's video says it will follow which hand someone writes with. Vercel's launch snippet is the brief: "Change the bottle cap to gold and keep everything else unchanged." ID: gpt_image_2_5_sunburst.
A/B with the same inputs
Do not change the prompt, the references, or the size when you compare Flare and Sunburst. If Flare is good enough, stay there. If the edit still crawls across the frame, Sunburst is the one that costs you wait time on purpose.
Scenes people are actually running
These are launch demos and day-one API examples, not a list of form fields.
Swap the jacket, leave the stance
OpenAI's developer post used this as the edit example: change a jacket and keep the subject in the same position. That is the test GPT-Image 2 often failed after two turns.
Rewrite one SKU detail on a pack shot
Upload the product photo. Name the one thing that moves (cap color, flavor banner, language on the label) and the list of things that must not. Sunburst is the starting ID; Flare is the check if you later need the same look at volume.
Poster, menu, or deck with real copy
The system card flags infographic layout. Put the headline in quotes. Check the spelling in the file you get back. 2.5 is better at this than 2.0; it is still not a typesetting app.
A room you can keep editing
The living-room walkthrough on the launch film is the point of 2.5: same architecture, new sofa, new daylight. Restate what is locked every turn. If a wall must stay pixel-identical, composite the approved patch onto the original.
Stay on 2, or leave the GPT-Image family
GPT-Image 2
Keep running production jobs here until a same-prompt A/B on Flare or Sunburst actually wins. Same 1K–4K range and reference-image budget.
Nano Banana Pro
Google's image line when the brief is Gemini, not GPT-Image.
Seedream 5.0 Pro
ByteDance stills, large reference sets, 1K and 2K.
Krea 2 Turbo
Cinematic text-to-image with no reference images.
GPT-Image 2.5 FAQ
Switching from GPT-Image 2, picking Flare or Sunburst, and what this page charges.
/ 01Should I replace GPT-Image 2 right now?
Only after an A/B with the same prompt and references. If 2 already ships the still, try Flare for wait time. If 2 drifts on edits, try Sunburst. GPT-Image 2 remains at /image/gpt-image-2.
/ 02Did OpenAI cut the price?
No. Flare and Sunburst use the same official per-million rates as GPT-Image 2. OpenAI's image generation guide now estimates 2.5 output tokens by model, quality, and size, and notes that the same quality name can use a different token count than GPT-Image 2. Credits on this page are in the table above; Flare and Sunburst are billed as separate IDs. Failed tasks are not charged.
/ 03Flare or Sunburst — which one?
Flare is the default, faster loop for drafts and volume. Sunburst is slower and meant for an edit that must leave the rest of the frame alone. There is no bare gpt-image-2.5 ID; this page uses gpt_image_2_5_flare and gpt_image_2_5_sunburst.
/ 04How do I run it on this page?
Use the playground at the top. Switch Flare and Sunburst in Model Type. Attach references only when the job is an edit or a composite.
/ 05Can I pick quality or a transparent background here?
Not yet. OpenAI documents extra quality steps (xhigh, max) and transparent backgrounds for 2.5. This playground does not expose those controls. Each request returns one image.