GPT Image in Batch Banana: best uses and how to prompt it

Best for

GPT Image is OpenAI's image model, and it takes your words literally. Give it the exact headline, the real numbers for a chart or the screen you want mocked up, and it puts them where you said, spelled as you wrote them. It is also careful in edits: change one thing and the rest stays put.

Sunburst is one of the slower models here. For painterly or cinematic images, try Nano Banana or Seedream. For vectors, use Arrow. Which model for what shows all five families on the same prompts.

A pixel-art phone level-select screen for Lantern Keep: a ruined castle at dusk, eight numbered stone nodes on a torch-lit path with three gold stars on levels 1 to 5, level 6 glowing, coins 1,240 and five hearts at the top, a green Play button at the bottom

A character turnaround of Fern the brown hare in an olive coat and a leather satchel, in front, three-quarter, side and back views at the same height, in ink and watercolour on off-white

Which version

Model	Best for	Credits per image	Reference images
GPT Image 2.5 Sunburst	OpenAI's most capable image model, built for precise edits	30	Up to 8
GPT Image 2.5 Flare	OpenAI's fastest model for everyday image generation	30	Up to 8

Both versions cost the same at every setting. Use Flare for new images; it takes about 16 seconds to Sunburst's 30. Use Sunburst for edits, where keeping the untouched parts exactly as they were matters most. On a fresh poster prompt, with the same reference poster attached for the look, both got every word right:

A stack of vinyl records in magenta and blue flash light, with LATE NIGHT RECORD FAIR set vertically down the right edge and the date and address small in the left corners

The record fair poster made on Sunburst: a close stack of vinyl in pink and blue light, LATE NIGHT RECORD FAIR down the right edge, the date and address small in the left corners

Write the prompt

Open with the subject and what the image is for. OpenAI: "Name the subject and intended use, such as a product photograph, advertisement, or diagram." For anything busy, write the prompt in labelled parts. OpenAI suggests "scene, subject, details, and constraints, using labeled sections". The brand board at the top lists its six panels one by one, in reading order.

Quote the text, say where it goes, and say it appears once. "Put required wording in quotes and describe its position and typography." Spell unusual names letter by letter. Then say that nothing else may appear: "Each line appears exactly once. No other text." Both posters below had the same reference poster attached for the look. The one whose prompt only mentions "the dates and address" got a date in 2025, an address in California and a tagline, all made up, and a lily where the records should be:

A neon-lit lily on a poster for the Late Night Record Fair, with invented dates in May 2025, an invented address in Riverdale, CA, and a tagline nobody asked for

To change one thing, say "change only" and list what stays. OpenAI: "say 'change only X' and list the details to preserve, such as identity, geometry, layout, lighting, or labels." Attach the image you're changing. On the Halvard board, a bare "Make it a night scene." took it literally: every panel went dark, and a moon, a lantern and a compass appeared. Naming the one panel to change, and saying the rest stays, redid the bottom-right photo and left the other five panels and every line of text alone:

Original

Make it a night scene

Change only one panel

When you edit the same image again, repeat the list. OpenAI's cookbook says to "repeat the preserve list on each iteration to reduce drift".

More tips

For slides and charts, give it the data. OpenAI's advice is to write "like an artifact spec rather than an illustration request" and to put the "numbers and labels directly in the prompt". The slide above was made from a prompt that listed every month's value.

Describe a screen as if the app already exists. "Describe the product as if it already exists. Focus on layout, hierarchy, spacing, and real interface elements." Name the buttons, the counters and what they read.

Give each reference a job. "Identify each input by number and purpose: subject, style, clothing, or background."

Keep a character fixed with a paragraph you repeat. Describe the character's defining details the same way in every prompt, and add "Do not redesign the character." The turnaround of Fern above attached Fern's picture and named what to keep: the ink and watercolour style, the coat, the satchel and the colours.

Write "photorealistic" when you want a photo. OpenAI's cookbook: include the word "to strongly engage the model's photorealistic mode". Lens and camera words help, but OpenAI calls them "cues for appearance, not a guarantee".

Settings

Quality is GPT Image's only setting besides the aspect ratio. It runs from Low to Max, and the default is High.

Model	Credits per image	Other settings
GPT Image 2.5 Sunburst	30	Quality: Low 3, Medium 12, Extra high 59, Max 118
GPT Image 2.5 Flare	30	Quality: Low 3, Medium 12, Extra high 59, Max 118

OpenAI's advice is to "compare quality levels before rewriting the prompt", and to use the top two levels "only when they improve an unmet quality requirement". Draft on Low, at 3 credits an image against 30 on High. In our test, Low got every number and label on this Halvard slide right; the higher levels added an axis and heavier type, not better facts:

A pitch-deck slide titled Halvard, first half of 2026, with a bar chart of bars sold per month from 1,200 to 3,100, June in orange, and three figures: 12,150 bars, 4.8 average rating, 41% buy again

The same Halvard sales slide at Medium quality, now with a labelled vertical axis behind the bars

GPT Image offers seven aspect ratios, from 16:9 to 9:16. It has no 21:9, 5:4 or 4:5.

Limits

It reads up to 8 reference images.
It doesn't make transparent backgrounds. Generate on a plain background and remove it afterwards.
Tiny text can still come out wrong. OpenAI suggests Medium or High for small text and dense panels.

Sources

OpenAI's own guides: Image prompting and the image prompting cookbook.

Images made in Batch Banana on October 2, 2026. Select Prompt on an image to read the prompt and model that made it. Side-by-side pairs and comparison rows show the first result.