OpenAI's newest image model. The same in-image typography as GPT Image 2 at half the credits, and the only model here that renders 4K in a portrait ratio — GPT Image 2 restricts its top tier to six wide ratios. Takes up to 16 reference images at no extra charge.
"A poster reading "[exact text]" in [font style], smaller line below reading "[subtitle]". [Paper and ink colours]. The text must be sharp and perfectly legible."
Quote both lines. The "sharp and perfectly legible" clause is what separates a clean render from a smudged one.
"A [package type] on [surface], printed label reading "[brand]" with smaller text "[details]" beneath, [light description]. Text must be crisp and readable."
Packaging is two text sizes at once. Naming both, with the surface and the light, is what holds the smaller line together.
"[Subject and scene], [light direction], shallow depth of field. — then set Resolution to 4K with a 3:4 or 9:16 ratio."
The reason to pick this model over GPT Image 2: its 4K tier works in portrait, where GPT Image 2 quietly drops to 2K.
Two things, in opposite directions. It renders 4K at any aspect ratio, where GPT Image 2 restricts 4K to six wide ratios and drops everything else to 2K — so portrait 4K only exists here. In exchange it drops three ratios (2:1, 1:2, 9:21) and will not take a pixel-size target, so masked edits and canvas expansion stay on GPT Image 2.
Both cost us the same upstream. GPT Image 2 is priced above that deliberately, from a period when its price was used to steer the auto-router toward it. This model is priced at the standard basis.
When the prompt's intent matches GPT Image 2.5's strengths and the cost is lower than alternatives that meet the same quality bar.
Yes. The model picker in any chat lets you pin per-turn or per-session. Set as default in /settings.
