GPT Image 2.5: OpenAI's New Image Model and How to Use It

What OpenAI Released

OpenAI released GPT Image 2.5 on 8 September 2026. It comes in two variants, and the names are worth knowing because you will see them mentioned elsewhere:

• Flare: OpenAI's default image model for most applications. Fast, with natural lighting and rich detail.
• Sunburst: a precision variant meant for premium work.

We run Flare, at high quality. That is the version this post is about.

What Changes Here

GPT Image 2.5 replaces GPT Image 2 everywhere the old model lived: Studio (create mode), the Agent and Fire Images. If you had a workflow built around GPT Image 2, it now runs on the new model without you changing anything.

The price stays at 8 credits per image, exactly what GPT Image 2 cost. The difference is that images are now generated at high quality instead of medium. Honestly, the same price for a better setting is the rare kind of update nobody has to think twice about.

Where It Is Strong

• Long, specific prompts. If your prompt reads like a brief with several requirements, this is the model built to follow it.
• Text and typography inside the image: product labels, posters, menus.
• Natural lighting. Scenes look lit, not rendered.

What It Does Not Do Here

Here, GPT Image 2.5 is a text to image model. You cannot upload a photo and ask it to swap the background. For editing uploaded photos, switch to Nano Banana 2.1 (7 credits per edit), Seedream 5 Lite (5 credits) or Flux 3 (10 credits).

It is also not the quick one here. Flare is billed as fast, but at high quality it still takes longer than the fast models, often tens of seconds per image. For rough drafts where you just want to see an idea, a cheaper model like Flux 3 [klein] at 2 credits is the better first stop. I'd bring GPT Image 2.5 in once the idea is settled and the details matter.

How to Use It in Studio

• Open Studio and stay in create mode.
• Pick GPT Image 2.5 from the model menu.
• Write the prompt (tips below).
• Choose an aspect ratio: 1:1, 16:9, 9:16, 4:3, 3:4, 2:3 or 21:9.
• Generate, then give it a little time.

If you prefer the Agent, describe what you want and name the model: "use GPT Image 2.5 for this". The Agent writes the prompt and shows a plan with the credit cost before anything is generated.

Prompt Tips That Actually Help

• Write the prompt like a brief: subject, style, composition, light. That order works fine.
• Put the exact text in quotes and say where it goes, for example the headline "Fresh Bread Daily" across the top of the poster.
• List the must haves as short sentences. "The cup is white. The table is oak. No people." beats one long sentence with ten commas.
• Pick the aspect ratio in Studio instead of describing it in the prompt.

A full example: A menu board for a small Italian café, chalkboard style, warm afternoon window light. Title "Pranzo" at the top in large hand lettering. Three items below it: "Lasagna 14", "Risotto 12", "Tiramisù 6". Plants on either side. No people.

Compare It in Fire Images

The honest way to know if GPT Image 2.5 is right for a job is to see it next to the others. Fire Images runs one prompt across several models at once and groups the results by model. Put GPT Image 2.5, Nano Banana 2.1 and Flux 3 on the same prompt, one image each, and the run costs 24 credits (8 + 6 + 10). For anything with text in it, that comparison is usually enough to decide.

Frequently Asked Questions

How many credits does GPT Image 2.5 cost?

8 credits per image, the same as GPT Image 2 cost, but now generated at high quality instead of medium.

Can GPT Image 2.5 edit my photos?

Not here. In Studio it works as a text to image model only. To edit an uploaded photo, use Nano Banana 2.1, Seedream 5 Lite or Flux 3 in Studio's edit mode.

Which variant runs here, Flare or Sunburst?

Flare, OpenAI's default image model for most applications, running at high quality.

Why does GPT Image 2.5 take longer than other models?

Because it runs at high quality, which takes longer than the fast models, often tens of seconds per image. If a generation fails, the credits are refunded automatically.