GPT-Image-2.5 Sunburst: Complex Layouts, Precise Edits
Get Started Free

GPT-Image-2.5 Sunburst

OpenAI's most capable image model for complex layouts, reference fidelity and iterative editing

Category
Image
Modality
Text → Image
Context
—
Released
Sep 2026
Run this model Estimate cost
Strengths

What it's the best tool for

  • OpenAI's most capable model for image generation and editing
  • Complex multi-layer layouts with predictable element placement
  • Up to 16 reference images, carrying a character or product across shots
  • True mask-based inpainting — the marked region changes, the frame stays
  • Natural lighting and rich textures for product and campaign imagery
  • Stays on track across a long series of consecutive edits
Limitations

When to reach for something else

  • No negative prompt: unwanted details are removed by rewording or a mask edit
  • No fixed seed, so an exact repeat of a frame is not guaranteed
  • Transparent background is preserved only in PNG and WEBP
  • The mask guides where the edit goes, it is not a hard pixel boundary
  • This is the top tier of the line; for speed and volume use GPT-Image-2.5 Flare
Sample output

How GPT-Image-2.5 Sunburst responds

Prompt
Editorial still life of a watchmaker's bench: a disassembled movement laid out in precise rows on a slate tray, brass tweezers, a loupe and a scatter of tiny screws under hard side light. Macro detail on brushed metal and jewel bearings, with no lettering anywhere in the frame.
GPT-Image-2.5 Sunburst
Generation time: ~35 sec
Where teams use it

Four scenarios where it pays for itself

01
Campaign creative
Hero visuals where every layer and every caption has to land
02
Product imagery
Goods with natural light, shadows and texture, without a studio
03
Consistent character
One hero across a series of shots, up to 16 references per request
04
Mask editing
Replace or remove a region on an existing shot without rebuilding it
About model

More about GPT-Image-2.5 Sunburst

GPT-Image-2.5 Sunburst: Complex Layouts, Precise Edits

GPT-Image-2.5 Sunburst is the top image model from OpenAI, released on 8 September 2026. It is built for work where generating a picture is not enough: multi-layer layouts, frames with a recognisable character, long sequences of edits on the same image. The model creates from a text description and edits existing shots, including replacing a single region through a mask.

Complex layouts and reference fidelity

The bet here is predictability. Elements land where the prompt puts them instead of drifting across the canvas, and lighting and textures read as photographed rather than drawn. Up to sixteen reference images can be supplied, so the same person or the same product carries from frame to frame without losing identity. Behaviour across consecutive edits deserves a separate mention: when one image goes through five rounds of refinement, it does not wander from step to step, which is exactly what production work usually lacks.

How mask editing works

This is real inpainting, not a full redraw. You mark the region with a mask and describe the change in words: swap the object, remove the clutter, add a detail. One nuance matters. The mask acts as a hint rather than a hard pixel boundary. The model treats it as guidance about where the change belongs and blends the result into its surroundings by light, scale and perspective instead of pasting a patch along the outline. For a complex scene that is an advantage; for a task that demands not a single pixel change outside the selection it is a limit worth knowing in advance.

Transparency, sizes and quality tiers

Background has three modes: auto, opaque and transparent. Transparency survives only in PNG and WEBP, since JPG cannot store an alpha channel. Dimensions are free-form, from 16 to 3840 pixels per side in steps of 16, so a tall banner and a square catalogue shot both come out without cropping. There are six quality tiers — auto, low, medium, high, xhigh and max — and on the top model the upper ones are the point. A single request returns up to twenty images, prompts run to 32000 characters, and two moderation levels are available.

Sunburst or Flare

Pick Sunburst when the output goes into a campaign, a product page or a design layout and has to survive close inspection. When volume and speed matter more and the idea is still being tested, the line also has GPT-Image-2.5 Flare, which is noticeably faster. Neither model supports a negative prompt or a fixed seed.

How to start

The model runs on NetRoom straight in the browser: no installation, no keys to configure. Write the prompt in plain language and download the result as JPG, PNG or WEBP. Billing is token-based, the cost depends on scene complexity and the chosen quality tier, and the current rate is shown in a separate block on this page.

Recent changes

What changed GPT-Image-2.5 Sunburst

  • + Added the GPT-Image-2.5 Sunburst (OpenAI) image model: complex layouts, up to 16 reference images, mask-based editing and transparent backgrounds.
Full changelog →

Use GPT-Image-2.5 Sunburst via the API

The same engine, straight from your code: one key and one balance for text, images, video and sound. Pay only for the requests you make.

curl
curl https://netroom.ai/api/v1/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "openai/gpt-image-25-sunburst", "input": {"prompt": "A cinematic mountain sunrise"}}'

The model id is already in the example. The full parameter reference and prices live in GET /api/v1/models and in the docs.

API documentation Get an API key

Try GPT-Image-2.5 Sunburst
right now

Free access to basic models. No card, no obligations.