Which model for what

Overview

Each model is good at different things. Some set text accurately, some are better at light and photographs, and two make vector files. Starting on the right model saves you credits and rewrites.

Start with the table of first picks below. Then compare five model families on the same six prompts, and see art from a game made with several models. Prices and the full list of models are in Models.

Start here

Job	First pick	Also good
Your product in a new scene	Nano Banana 2	GPT Image 2.5 Sunburst
The same character in new scenes	GPT Image 2.5 Sunburst	Nano Banana 2
Exact text, slides and infographics	GPT Image 2.5	Nano Banana Pro
Posters, logos and packaging	Ideogram 4.5	GPT Image 2.5
Cinematic photos and portraits	Seedream 4.5	Nano Banana Pro
Placing things exactly on the canvas	Seedream 5 Pro	Ideogram 4.5
Icons, marks, anything you'll scale	Arrow 2	Recraft V4.1 SVG
A real place or product, by name	Flux 3 Image with Web search	None; no other model searches the web
Cheap drafts	GPT Image 2.5 on Low quality	Nano Banana 2 Lite

These picks are our judgement. They come from what each maker's guide says its model is for, and from one run of the comparison below, not from a benchmark. Your prompt may disagree with them. When you're unsure, add two or three models to a variety pack and run the prompt once on each.

What every model agrees on

These four rules come up again and again in the makers' guides, and they work on every model here.

Quote the exact words. Any text that must appear goes in quotes, with where it sits and how it looks.
Say what the image is for. A logo for a small chocolate maker, a slide for a pitch, a sticker for a shop. The purpose guides the model's choices of style, layout and detail.
Give each reference a job. Say what to take from each image you attach: the character, the palette, the background.
When you change an image, say what changes and what stays. List the parts that must not move.

Each guide has the details for its model, with examples.

Same prompt, five models

We ran six prompts and made two images on each family. Nano Banana runs on Nano Banana 2 and Seedream on 5 Pro, both at 2K; GPT Image runs on Sunburst and Ideogram on 4.5, both on High. We show the first image of the two, or the second where we liked it more. Nothing is retouched.

Product shot

Halvard bar on an open atlas of Norway by a window, HALVARD large on the cobalt wrapper and a small added DARK 72% label

GPT Image put a stray dot before the D in "HALVARD"; the other four set it clean. Nano Banana 2 and GPT Image both added small print nobody asked for.

Game asset

This row is not one prompt on five models. It is art from Geolingo, a geography game we built, and every asset in the game was made in Batch Banana.

Game icon of Big Ben in warm stone with a dark slate spire and two clock faces, on plain cream

White quiz card on a stack of red-edged cards, a faint map border around the text Starts With C: North America, 3 countries in North America start with C, and a 3 countries pill

Nano Banana 2 made about 2,100 of the project's 2,800 files, including every landmark icon. GPT Image 2 made icon sheets, treasure chests and paywall drafts. Krea 2 and Nano Banana Pro made card backs, and Ideogram 4 a region wordmark. Each asset used earlier Geolingo art as a style reference, which this page does not show. The Big Ben prompt is cut to its subject; the full prompt attached nine style images.

Poster with text

This row did not use one shared prompt. We wrote one brief, and Agent adapted it into a prompt for each model, so each image opens its own prompt. Every prompt mentions a style reference, Image 1, which this page does not show.

Fresco poster with MORE in tall condensed white letters over slate-blue clouds, BATCH BANANA MODEL SHOWCASE on one line just above it, figures crowding the lower half

Fresco poster with MORE in a white serif crossed by a hairline with small stars, two pairs of figures reaching toward each other above and below it

All four spelled every line right, and all four put MORE across the middle. GPT Image, Ideogram and Seedream set BATCH BANANA and MODEL SHOWCASE in the top corners. Nano Banana 2 ran them together as one line just above MORE. GPT Image and Seedream drew nearly the same headline: a tall serif crossed by a hairline with small four-pointed stars. GPT Image arranged the fresco as two pairs of figures reaching toward each other across the word. Ideogram filled the whole sheet with figures on grey clouds under a gold sky. Nano Banana 2 and Seedream kept open blue sky around the word and put most figures below it. The infographic with labels, at the end of this page, is a harder test of text.

Character from a reference

Watercolor of Fern on a crate at a lantern-lit night market reading a You Are Here map, badgers and an owl shopping behind

Painterly scene of Fern perched on a wooden crate under a paper lantern at a night market, reading a map

Each model got the same reference, Fern the hare in an olive coat, made for these guides. All five kept the coat and the satchel. Nano Banana 2 filled the market with badgers and an owl.