ShimmerPic
GUIDE

How to Use GPT Image 2.5: Sizes, Quality, Reference Photos

A practical guide to GPT Image 2.5 in the browser: the two models, six aspect ratios, 1K to 4K, Standard vs High, reference photos, prompt limits and cost.

How to Use GPT Image 2.5: Sizes, Quality, Reference Photos

GPT Image 2.5 is OpenAI's image model released on September 8, 2026. In the API it comes as two models, Flare and Sunburst, and ShimmerPic runs both. This is the guide to every setting you will see in the Studio and what it does to the result and the price.

1. Pick a model

  • Flare — the fast default. Same quality and editing improvements as the whole 2.5 generation, at 50% lower latency than GPT Image 2.
  • Sunburst — the precision model, with longer generation times, meant for polished product and campaign work.

Both cost the same credits here. If you are unsure, start on Flare and switch for the final render. The side-by-side comparison shows the difference on real prompts.

2. Choose an aspect ratio

Six ratios are available: 1:1, 4:5, 3:2, 2:3, 16:9, 9:16. Pick the one your destination needs rather than cropping afterwards — the model composes for the frame you give it. 4:5 and 9:16 are the social-feed and story sizes; 3:2 and 16:9 suit web heroes and slides.

3. Choose resolution and quality

Resolution is 1K, 2K or 4K; quality is Standard or High. Together they set the price:

Credits per image1K2K4K
Standard3612
High122440

A reasonable habit: explore at 1K Standard (3 credits), then re-run the keeper at the size you actually need. The pixel dimensions at each tier vary by ratio; not every ratio reaches 4,096 px on the long side at 4K.

4. Add reference photos if the prompt needs them

Up to three reference images per generation, each up to 10 MB and 4,096 px on the long side. Each one adds 3 credits. GPT Image 2.5 is noticeably better than its predecessor at keeping the subject of a reference recognisable across new settings and styles, which is what makes reference-led prompts — "this product, in that scene" — practical.

Prompts in the library that need a reference are labelled Needs a reference photo on their cards.

5. Write, or borrow, the prompt

Prompts can be up to 3,000 characters (6,000 UTF-8 bytes). The first 500 bytes are included in the base price; each additional 1,000 bytes adds 2 credits, so very long prompts cost a little more. The Studio warns before you generate if a prompt is over the limit rather than silently cutting it.

If you would rather not start from blank, the prompt library has 531 community prompts with their real output, grouped by use case — product photography, posters, characters and more. Open one, change the bracketed parts, and generate.

6. Generate, then edit

Every result is saved privately to your creations. From there you can download it, or continue editing: open the image as a reference and describe the change. GPT Image 2.5 was built to follow "editing instructions more reliably across multiple turns", so iterative edits — change the background, keep everything else — are where it improved most.

What it costs to start

New accounts get 6 free credits: two 1K Standard images. After that, credits come in one-time packs from $9 or in monthly and annual plans. There is no unlimited tier. The full breakdown, including cost per image at each setting, is in What GPT Image 2.5 costs per image.

Source for model facts: OpenAI, Introducing ChatGPT Images 2.5. ShimmerPic is an independent tool and is not affiliated with OpenAI.