GPT Image 2.5 AI Image Generation

GPT Image 2.5 is an OpenAI image generation model and one of the strongest options on FlyAIgh for any image where readable text is part of the deliverable — posters, ad creatives, product labels, slide graphics, infographics. Where most image models render text as noisy lookalike glyphs, it produces clean letterforms that match what you typed.

It ships in two lines that cost exactly the same. **Sunburst** is the finest line: the one to pick when the lettering has to be exact, the layout is complex, or you are rendering at the Max quality tier. **Flare** is the fast everyday line for quick, stable edits. They differ in model behaviour, not in billing — there is no premium for choosing Sunburst.

Both lines take five quality tiers (Low, Medium, High, Extra high, Max), render at 1K, 2K or 4K, and support 15 aspect ratios from 1:1 through 21:9 and 1:3. Both do text-to-image and image editing, accepting up to 16 reference images in edit mode, and both can output a transparent background — in text-to-image as well as in edits — for logos, stickers and product cut-outs.

Pricing is token-based, so the credit cost moves with what you ask for rather than sitting at a flat rate. At Low or Medium quality a 1K image is 1 credit; High at 2K is 7; Max at 4K at 1:1 is 40. Each reference image adds roughly half a credit. **One effect worth knowing: aspect ratio changes the price**, because a square image spends the full pixel budget and a wide one does not — at Max 4K, 1:1 costs 40 credits while 21:9 costs 14. Symmetric ratios cost the same (16:9 and 9:16 are identical). The exact total is always shown before you generate.

GPT Image 2.6, the newer generation, has its own page. It adds a free Complexity control and costs more than 2.5 at Low and Medium quality; from High up the two cost about the same on square images, and 2.5 stays cheaper on wide ones and keeps the five extreme ratios and transparent text-to-image. GPT Image 2, the earlier generation, also has its own page.

The model is particularly strong at faces, hands, on-image text, structured layouts and any prompt with explicit lettering ('a poster that reads ...'). It is weaker than Midjourney for distinctive artistic style. On FlyAIgh it runs on the same credit wallet as every other model — no separate OpenAI subscription or API key.