Which AI model generates true 4K images, and is it worth it?

Detail density, and very little else. The model composes the same scene at every resolution — the same pose, the same framing, the same interpretation of your prompt. What changes at 4K is fabric weave, skin texture, the edge of a leaf, and whether a printed line of text survives.

That has a direct workflow consequence: composition questions are cheap to answer. Find the shot at 512 px for $0.03, then regenerate the keeper at 4K. Paying $0.20 to discover that the pose is wrong is the most common way to waste money here.

Three cases. Text in the image — packaging, a poster, a UI mock-up, a sign: Pro is the model that renders letterforms legibly instead of approximating them. Print delivery, where 4K is a hard requirement rather than a preference. Identity-critical work, because Pro's typed reference budget (up to 6 objects and 5 characters) holds a face or a product tighter than generic references do.

Outside those, Nano Banana 2 is usually the better buy. Its 2K output at $0.09 is enough for any screen use, and its 512 px tier makes exploration almost free.

Veo 3.1 and Veo 3.1 Fast render 4K; Veo 3.1 Lite does not, and Gemini Omni Flash is 720p only. The price step is steep — an 8-second 4K clip with audio is $4.40 on Veo 3.1 against $3.00 at 1080p.

There is a practical argument for skipping it beyond cost: 4K is where capacity errors concentrate, especially in combination with a last frame. On Veo 3.1 and Veo 3.1 Fast, 1080p is priced identically to 720p and queues far less, which makes it the sensible default for anything that is not a 4K deliverable.