GPT Image 2.5 — Faster, Sharper, Built for Editing

The direct successor to GPT Image 2, shipped as two models: Flare for speed and Sunburst for quality. On this page you use it through the generator at the top: the Version switch picks Flare or Sunburst, the Generate tab works from a text prompt, and the Edit tab takes up to 16 reference images plus an optional mask. Results can be downloaded as PNG, JPEG or WebP.

Flare is the speed-oriented model with quality roughly comparable to GPT Image 2. Sunburst is the quality-oriented model with better editing precision and fidelity. Both cost the same credits. Draft on Flare, move to Sunburst only when Flare falls short, then drop back to Flare once the prompt is dialed in.

Same lineage, refined: a speed-oriented Flare model, more natural light and texture, better reference-image preservation, more stable multi-turn edits, and new xhigh / max quality tiers. The raw-quality gain is modest — the editing and consistency improvements are the real upgrade. On imagesv2 the pricing also works differently. GPT Image 2 charges a fixed price per quality tier at any size, while GPT Image 2.5 scales with the number of output pixels. A low-quality 1024×1024 draft costs 10 credits on either model, but a high-quality 2048×2048 image costs 110 credits on GPT Image 2 and 220 on GPT Image 2.5. If you need one large, high-quality image and no editing, GPT Image 2 on the homepage can be the cheaper option.

No. You do not need an OpenAI account or an API key; imagesv2 gives you GPT Image 2.5 in the browser with credits. New accounts get 3 credits, and the first image at low quality costs exactly 3 credits on either Flare or Sunburst, including image-to-image with reference images. Medium and higher tiers unlock after your first purchase, and images made before a purchase carry a watermark until you buy any credit pack or plan. No credit card is needed to sign up.

In credits, per image, and Flare and Sunburst cost the same. Unlike GPT Image 2, the price of GPT Image 2.5 grows with the output size, because generation cost scales with pixel count. At 1024×1024 an image costs 10 credits at low, 60 at high, 100 at xhigh and 220 at max. At 2048×2048, four times the pixels, the same tiers cost 30, 220, 380 and 850. Each reference image in edit mode adds 20 credits, and generating 4 images multiplies the total by 4. The Generate button shows the exact figure for your size, tier and references before you click, so set the size first and watch the number. See the plans above.

Up to 16 reference images per edit request, with the mask applied to the first. Four references are usually enough for identity locking. Each reference adds 20 credits to the image price, so 4 references add 80 credits and 16 add 320. Add a reference only when it shows something the others do not, such as a different angle, the logo, or the color palette.

Yes. Pick transparent background with PNG or WebP output for a real RGBA file. Sizes can be 1024×1024, 1536×1024, 1024×1536, or a custom WIDTHxHEIGHT (multiples of 16, aspect ratio up to 3:1, up to 3840 px on a side). On GPT Image 2.5 the model produces transparency itself, so there is no separate cut-out step and no extra charge; on GPT Image 2 a transparent background adds 10 credits for background removal. JPEG cannot store transparency, which is why Transparent requires PNG or WebP. Custom sizes must also stay between 655,360 and 8,294,400 total pixels; 3840×2160 is the largest 16:9 size.