GPT Image 2.5 Image Generator
Describe a picture, or add up to 3 photos to edit. 1K costs 4 credits, 2K costs 6 and 4K costs 9.
- Released:
- 8 September 2026, by OpenAI
- Output:
- 1K, 2K or 4K, seven aspect ratios
- Version:
- Flare
What GPT Image 2.5 made in our tests
We ran GPT Image 2.5 on 16 October 2026 with seven prompts: four built around lettering, one photo edit, one 2K product shot and one street photo. Each ran once. Press Try this prompt to load it into the form above, or copy it.
A chalkboard menu with four prices
3:4, 1K, text to image, 74 s
Every item and every price came out exactly as typed. It added dotted leaders between names and prices, which a real sign painter would do too.
Prompt we used
A cafe chalkboard menu hanging on a brick wall, hand-drawn chalk lettering. Title: "TODAY". Four lines, each with a price: "Flat white 4.50", "Cardamom bun 3.80", "Iced matcha 5.20", "Soup of the day 7.00". Small chalk doodle of a coffee cup in the corner. Even daylight, straight-on view.
A three-panel comic with one character
16:9, 1K, text to image, 42 s
Three panels in order, both captions and the speech bubble spelled right, and the same robot in every panel. The fastest render of the batch.
Prompt we used
A three-panel comic strip in a clean flat style, laid out left to right. Panel 1: a small orange robot waters a single plant on a windowsill, caption "Day 1". Panel 2: the plant is now taller than the robot, caption "Day 30". Panel 3: the robot sits in the shade of a huge leafy plant reading a book, speech bubble "Worth it." Consistent robot design in all three panels, white gutters between panels.
A letterpress wedding invitation
4:3, 1K, text to image, 69 s
All four lines in the right order with a believable pressed-ink look. It added a small leaf ornament between the lines, nothing written.
Prompt we used
A flat-lay photo of a letterpress wedding invitation on cotton paper with dried eucalyptus around it. The card reads, centred: "Maya & Theo", then "are getting married", then "Saturday 14 June", then "The Old Boathouse". Elegant serif type, sage green ink, soft natural light from the left.
The night market poster
3:4, 1K, text to image, 85 s
All three quoted lines are spelled as written. It added one unrequested character, a noodle sign on a stall banner.
Prompt we used
A vertical poster for a neighbourhood night market. Headline in bold hand-lettered type: "NIGHT MARKET". Subline: "Fri & Sat, 6pm till late". Small print at the bottom: "Lanterns, noodles, live music". Warm lantern light, paper-cut illustration style, deep teal and amber palette.
Photo edit: same person, rainy street
1:1, 1K, 1 reference photo, 95 s
Face, glasses, curls, hair colour, the dark top and the apron all held. It invented a readable COFFEE neon sign and gave her a tote bag.
Prompt we used
Keep the same person: same face, curly auburn hair, round glasses and cream apron. Move her outside the cafe on a rainy evening street, holding a paper cup, neon sign reflections on the wet pavement.
A 2K product shot with a printed card
1:1, 2K, text to image, 47 s (shown resized)
The card reads SMALL BATCH exactly and the file is 2048 by 2048 pixels. Look closely: the jug has a see-through window no real ceramic jug has.
Prompt we used
Product photo of a ceramic pour-over coffee set on a pale oak table, morning light, soft shadows, a small card beside it reading "SMALL BATCH" in clean sans-serif type.
A vertical street photo, later animated
9:16, 1K, text to image, 74 s
Convincing light and shadow. We used it as the first frame of a MiniMax H3 clip, linked below.
Prompt we used
A red vintage bicycle with a wicker basket of lemons leaning against a pastel yellow wall on a narrow cobblestone street, morning sun, long soft shadows, potted plants by a blue door, photographed at eye level, realistic photo.
Next: animate a still with the MiniMax H3 video generator.
GPT Image 2.5 versus Nano Banana 2.1 on the same prompts
Three of our prompts are the exact ones on our Nano Banana 2.1 page, so the two models can be compared on identical inputs. One run each: a hint, not a verdict.
GPT Image 2.5
Nano Banana 2.1GPT Image 2.5
Nano Banana 2.1| Test | GPT Image 2.5 | Nano Banana 2.1 | Pick |
|---|---|---|---|
| Poster, three quoted lines | All three lines correct, plus one noodle character on a banner. 85 s. | All three lines correct, no extra words. 44 s. | Nano Banana 2.1 when only your words may appear. |
| Edit: same person on a rainy street | Hair colour held. Invented a COFFEE neon sign. 95 s. | Hair came back brighter red. No readable invented text. 48 s. | GPT Image 2.5 when the person must match; check for stray signs. |
| Product shot at 2K | SMALL BATCH exact; odd see-through jug. 47 s. | SMALL BATCH exact; plausible objects. 53 s. | A draw on text; check objects closely on both. |
GPT Image 2.5 Flare versus Sunburst
OpenAI released GPT Image 2.5 as two versions on the same day. Our renders above are all Flare.
| Compared on | Flare | Sunburst |
|---|---|---|
| Built for | Speed: the default for most work | Quality and edit precision |
| Speed | Faster; 42 to 95 s per image in our run | Slower, per OpenAI |
| Edits across turns | Good: our edit kept face and hair colour | Described as better at holding earlier edits |
| On CastTake | Yes, this is the version we sell | Not offered yet |
GPT Image 2.5 price, and which image model to pick
Credits per image, from the same rate card the Generate button uses. All three are in the picker above.
| Model | Credits per image | Best for |
|---|---|---|
| GPT Image 2.5 | 1K 4, 2K 6, 4K 9 | Menus, comics and cards with several lines of text; edits that keep a face; 4K files. |
| Nano Banana 2.1 | 1K 3, 2K 4 | The cheapest clean result when nothing beyond your quoted words may appear. |
| Nano Banana Pro | 1K 10 to 4K 13 | The priciest image model we sell, for when the others miss. |
Every model and price is on the models page; credits come from a plan or a top-up pack.
How to use GPT Image 2.5 on CastTake
- GPT Image 2.5 is already selected in the form above. Pick an aspect ratio and a size: 1K, 2K or 4K.
- Write the prompt. Put every word that must appear inside quotation marks, in reading order.
- Optional: add up to 3 photos to edit instead of starting from nothing.
- Check the credit price on the Generate button, press it, and the image appears beside the form in about a minute.
Limits we hit with GPT Image 2.5
What went sideways, so you know what to check before you rely on a result.
- It adds small things of its own: a noodle character on the poster, a COFFEE sign in the edit, a tote bag nobody asked for.
- Objects can be subtly impossible. The 2K pour-over jug has a see-through window in solid ceramic.
- It is not quick: 42 to 95 seconds per image in our run, slower than Nano Banana 2.1 on two of three matched prompts.
- Each prompt ran once. Treat the comparison as a hint and test your own prompt; a 1K image is only a few credits.
| Image | Size | Inputs | Render time |
|---|---|---|---|
| Night market poster | 3:4 at 1K | Text | 85 s |
| Pour-over product shot | 1:1 at 2K | Text | 47 s |
| Rainy street edit | 1:1 at 1K | 1 photo | 95 s |
| Chalkboard menu | 3:4 at 1K | Text | 74 s |
| Robot comic | 16:9 at 1K | Text | 42 s |
| Wedding invitation | 4:3 at 1K | Text | 69 s |
| Bicycle street | 9:16 at 1K | Text | 74 s |
GPT Image 2.5 questions
What is GPT Image 2.5?
GPT Image 2.5 is the image model OpenAI released on 8 September 2026, the successor to GPT Image 2. It makes pictures from a text description and edits photos you give it, and it is especially good at writing text inside images. On CastTake you pick it in the model list above, type a prompt, optionally add up to 3 reference photos, and see the credit price before you press Generate. We ran seven prompts on 16 October 2026 and published every result with its prompt and render time on this page.
What is the difference between GPT Image 2.5 Flare and Sunburst?
OpenAI shipped GPT Image 2.5 as two versions. Flare is the speed-first default, which OpenAI pitches as faster than GPT Image 2 at higher quality. Sunburst is the slower version tuned for fine detail and for holding earlier edits and reference subjects across several turns. CastTake sells Flare. In our run it took 42 to 95 seconds per image, kept a face and hair colour through an edit, and spelled every quoted line correctly. If you need many rounds of edits on one image, check results closely and run a second try when a detail drifts.
Is GPT Image 2.5 free?
There is no free tier for it on CastTake, but it is cheap to try. One image costs 4 credits at 1K, 6 at 2K and 9 at 4K, and the form shows the total on the Generate button before you press it, so a handful of 1K tests costs less than a single video clip. Every prompt on this page can be copied, which saves a few retries. ChatGPT also offers GPT Image 2.5 inside its own plans, with usage limits that depend on the plan you are on.
Is there a GPT Image 2.5 API?
Yes, OpenAI sells API access to both Flare and Sunburst for developers who want to build image generation into their own app, with keys, rate limits and per-image pricing set by OpenAI. CastTake does not resell the API. What we offer is the finished tool: type a prompt, add photos if you want an edit, see the price, and download the PNG. If you only need images and not an integration, that saves you writing code, handling keys and building an upload flow, and every image stays in your history to download again later.
GPT Image 2.5 or Nano Banana: which is better?
It depends on the job, and our matched runs show it. On the same poster prompt both spelled every line, but GPT Image 2.5 added one stray character and Nano Banana 2.1 added none. In the photo edit GPT Image 2.5 kept the hair colour while Nano Banana 2.1 drifted, though GPT invented a neon sign. Nano Banana 2.1 is cheaper at 3 credits for 1K against 4, and it stops at 2K while GPT Image 2.5 goes to 4K. Use GPT for text-heavy layouts and faithful edits, Nano Banana 2.1 for the cleanest cheap result.
Can GPT Image 2.5 write text inside images?
It was the strongest thing we saw. A chalkboard menu with four items and four prices, a wedding invitation with four lines and a comic with two captions and a speech bubble all came out spelled exactly as typed, on the first try. Put the exact words in quotation marks, list them in the order you want, and say where each should sit. It sometimes adds decoration of its own, such as dotted leaders, ornaments or a sign in the background, so read the whole image before you publish it, not only the lines you asked for.
More from CastTake
- MiniMax H3 video generatorTurn a GPT Image 2.5 still into a clip with sound. Our bicycle photo became one.
- Nano Banana 2.1 guideThe cheaper image model, with its own test renders.
- Every model and its priceThe full rate card for every image and video model we sell.
- AI image generatorThe same form with every image model in one picker.