GPT Image 2.5 Review: Tested with Real Prompts
GPT Image 2.5 review with real prompts run on GPT-IMG: where it beats GPT Image 2, where it still falls short, and what each result costs in credits.
GPT Image 2.5 is OpenAI's newest image model, released September 8, 2026. Two API variants, Flare and Sunburst, both bill at the same token price as GPT Image 2. On GPT-IMG it follows dense instructions in a single pass, renders multi-word in-image text correctly, and holds both the subject and the crop through a reference edit.
Created: September 9, 2026. Updated: September 9, 2026.
What GPT Image 2.5 actually is
GPT Image 2.5 arrived on September 8, 2026, and OpenAI's announcement frames it as an update rather than a generational leap. The headline change is not a new art filter. It targets the parts of production that cost the most time: instruction following, edit containment, iteration speed, and reference-image consistency. Generation latency drops by up to 50% compared with the GPT Image 2 model OpenAI released five months earlier.
For developers, OpenAI splits the new release into two API routes. GPT-Image-2.5 Flare is the default for most apps: social content, product imagery, rapid prototyping, and high-volume generation. GPT-Image-2.5 Sunburst is the slower, more careful variant aimed at production campaign creative and detailed product work. Both bill at the same per-token rate, so cost is not a reason to pick one over the other. It is a job-fit decision.
On GPT-IMG, only the Flare route ships today, so the rest of this GPT Image 2.5 review covers the version our prompt box actually sends to. The prompts below behave the same way whether you run them through ChatGPT or through the GPT Image 2.5 page.
Where it pulls ahead of GPT Image 2
Instruction following
The baseline brief below stacks six concrete constraints in one prompt: subject, handle orientation, light direction, background tone, a reserved empty area, and a no-text rule. The cover image at the top of this post is what came back, with all six intact in a single pass.
A square studio product photo of a cobalt-blue ceramic mug,
handle on the right, soft light from the left, pale gray
background, leave upper-left empty for headline, no text.
Run: 1:1 / 1K / Standard, 10 credits.
If you want tighter control over the prompt structure itself, the text-to-image prompt guide walks through the same role-based format this brief uses.
In-image text
Short headlines were never the hard part. This one landed exactly as written, centred, with the margins the prompt asked for:
A 16:9 event poster with a single headline reading
"SPRING LAUNCH" in clean sans-serif, centred on a deep
green background, generous margins, no other text.
Run: 16:9 / 2K / Medium, 20 credits.
The interesting test is six multi-word labels in one frame, which is the shape of brief that has historically come back with letters missing:
An infographic with six labeled regions in a row:
"SIGN UP", "VERIFY EMAIL", "UPLOAD REFERENCE",
"GENERATE", "REFINE", "DOWNLOAD", clean sans-serif,
light background.
Run: 16:9 / 2K / Medium, 20 credits.
All six labels are spelled correctly. So are the six captions the model added on its own initiative, apostrophe included. The icons match their steps, the arrows point the right way, and the numbered badges run 1 to 6 in order.
That is one run, not a guarantee. In-image text is still the part of any image model most likely to break, and the sensible habit for anything legally or commercially load-bearing is to read every word before you ship it. But the old advice for this model generation — keep body copy out of the image and composite it in afterwards — did not match what came back here.
Reference edits
Uploading the cover image and asking for one change:
Keep the cobalt-blue mug exactly as it is, same angle and
same lighting. Replace the pale grey background with a
warm terracotta wall.
Run: 16:9 / 1K / Standard, 25 credits.
The handle, the angle, the glaze highlights, and the gradient down the ceramic all survive. So does the framing, which is worth noting: running the same kind of background swap on Nano Banana Pro returned a subject that had drifted larger and towards the centre despite an explicit instruction to hold the crop. On this run, GPT Image 2.5 held it.
Where it stumbles
Nothing in this round produced a broken image, so the honest limitations are about cost and coverage rather than quality.
The top of the price range is steep. A 4K High render costs 120 credits, and a 4K High reference edit costs 130. That is one image for more than four times the price of a 1K Standard draft, so the drafting-cheap-then-finishing-expensive habit matters more on this model than on a flat-rate one.
Reference edits jump to a higher table. A 1K Standard generation costs 10 credits; the same job with a reference photo costs 25. Sessions built on repeated reference edits add up faster than the headline price suggests.
Long edit chains were not tested here. Every edit in this review was a single round on a fresh image. Multi-turn behaviour, where a subject drifts across the fourth or fifth consecutive revision, is a different question and this post does not answer it. Until it does, keep an editing session short and return to a saved version rather than correcting forward through a long chain.
For more on controlling this kind of edit, the reference image workflow guide covers the prompt patterns that hold up across multiple rounds.
Parameters and controls in GPT-IMG
| Setting | Values |
|---|---|
| Resolutions | 1K, 2K, 4K |
| Quality tiers | Standard, Medium, High |
| Aspect ratios | 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9 |
| Reference images per generation | Up to 5 |
Quality and resolution are independent. A 1K Standard run and a 4K High run both produce a finished image, but at very different price points and detail levels. Reference images accept up to five photos per generation. There is no separate role selector: the prompt is what tells the model which photo is the subject and which one is only there for style.
Aspect ratio is set before the prompt is written, not after cropping. A 16:9 brief and a 9:16 brief are genuinely different compositions.
Credit cost on GPT-IMG
It shares its credit table with GPT Image 2 on GPT-IMG. The price is in credits, not dollars. One credit pack sells at $4.99 for 300 credits, so the dollar conversion is a moving target. Use the table below to size the work you are about to do, and check the pricing page for the latest dollar numbers.
| Resolution | Standard | Medium | High |
|---|---|---|---|
| 1K | 10 | 20 | 65 |
| 2K | 15 | 20 | 65 |
| 4K | 25 | 30 | 120 |
Reference-image edits bill on a separate table from text-to-image, because the model reads a source photo instead of generating from a blank slate. A 1K Standard reference edit costs 25 credits. A 4K High reference edit costs 130 credits.
The cheapest draft is 10 credits. The most expensive single image is 130 credits. A practical workflow is to draft at 1K Standard until the prompt and the composition are right, then move to 2K or 4K for the final render. Failed jobs refund credits automatically, so iteration does not punish the budget.
How to run GPT Image 2.5 on GPT-IMG
- Sign in or create an account. New accounts get 30 credits, enough to run a 1K Standard draft and a 2K Medium follow-up.
- Open the GPT Image 2.5 page. The generator box is pre-set to the model, 1:1, 1K, Standard, the cheapest draft setting.
- Paste a prompt, or pick one of the templates below the box. Set the aspect ratio to match the placement: 9:16 for a story, 16:9 for a hero, 1:1 for a feed post. If you are editing an existing photo, click the reference image slot and upload.
- Generate. The credit cost shows on the button before you submit. Failed jobs return credits to your balance.
FAQ
What is GPT Image 2.5?
The September 2026 image model from OpenAI is available in ChatGPT, ChatGPT Work, and Codex, with Flare for speed and Sunburst for tight edit control both at the same token rate. On GPT-IMG it is pre-selected on its own model page and on GPT Image prompts opened from the gallery.
How is GPT Image 2.5 different from GPT Image 2?
It follows dense instructions more reliably and preserves reference photo identity better than GPT Image 2, and OpenAI puts generation latency up to 50% lower. On GPT-IMG the credit cost is identical, so the choice is about quality, not budget. Multi-turn edit behaviour is not something this review tested.
How much does GPT Image 2.5 cost per image on GPT-IMG?
A 1K Standard draft costs 10 credits and a 4K High render costs 120. Reference-image edits run on a higher table, from 25 credits at 1K Standard to 130 at 4K High. The full grid is in the credit section above, and failed jobs refund automatically.
Can GPT Image 2.5 edit existing images, or only generate from text?
Both. Upload a reference image and describe the change. In the test above it made only the change asked for, keeping the subject, the lighting and the crop. Reference edits bill on a higher table than text-to-image, so a 1K Standard edit costs 25 credits rather than 10.
That is the GPT Image 2.5 review, run on the version GPT-IMG uses today: six-constraint briefs landing in one pass, six multi-word labels spelled correctly, and a reference edit that moved the background without moving the mug. The cheapest way to find out where it fits your own brief is to run a real prompt on the GPT Image 2.5 page.