Model comparison articles go stale the week after they are published. Running the comparison yourself on your own prompt does not. The image composer takes a set of ticked models rather than one, fans a single press out across all of them in parallel, and groups everything that comes back as one album. The thing to understand before you press it is which wallet each row spends from.
The model picker is a set of checkboxes, not a dropdown. Whatever is ticked when you press Generate gets the same prompt, the same reference images, the same aspect ratio and the same seed policy, and every one of them is submitted at once. They finish at different times because they run on different hardware — our GPU serialises its queue, the hosted providers do not — so the album fills in over a minute or two rather than all at once.
Grouping matters more than it sounds. Every job from one press shares an album id, so the gallery shows them as one run rather than as eight loose cards you have to remember the order of. Days later, the album is still the record of "this prompt, these models, this day", which is the thing that makes the comparison worth having done.
This is the part worth reading twice. There are two wallets and they are not interchangeable. ⚡ green refills on a clock and can only pay for models running on hardware we own. ✦ gold is bought with money and pays for anything, including every hosted model. A press with three own-GPU models and two hosted ones ticked spends from both wallets in the same submit, and the button quotes the split before you commit to it.
| Model | Runs on | Wallet | Price |
|---|---|---|---|
| Z-Image Turbo | Our own GPU | ⚡ green | from ⚡1 |
| Qwen Image Edit (Local) | Our own GPU | ⚡ green | from ⚡1 |
| Krea 2 Turbo | Our own GPU | ⚡ green | from ⚡1 |
| GPT Image 2 | A subscription we own | Neither | 0 credits, day-limited |
| Seedream 4.5 | Wavespeed | ✦ gold | ✦1.5 |
| Nano Banana Pro | Wavespeed | ✦ gold | ✦4.7, ✦8.1 at 4K |
What it costs: Own-GPU prices start at ⚡1 an image, are set at runtime, can carry a demand multiplier when the queue is deep, and fall back to a gold-billed API lane when the local hardware is unavailable. The number on the Generate button is the quote for that press; ⚡1 is the floor, not a promise.
A free account has no gold at all — signing up grants none — so its picker fans out across the own-GPU rows plus GPT Image 2, and the hosted rows show the plan that opens them instead of a price. That is still a real comparison of four or five distinct models, which is more than most people run before choosing one.
Which plan: Any signed-in account can fan a prompt out across models. What differs by plan is which rows are tickable: a free account reaches the own-GPU models and GPT Image 2, and every hosted model opens from Starter and is paid in ✦ gold. Locked rows stay visible in the menu with the plan that opens them, so the picker doubles as the price list.
One habit worth forming: run the comparison once at the start of a project and then stop. Fanning every prompt across six models for weeks is an expensive way to avoid making a decision, and the album from the first run is usually still the right answer a month later.
Load a comparison prompt — Fills the prompt bar. Tick the models you want and press Generate yourself.
Models declare what inputs they need, and a press only submits the ones whose needs are met. An edit model that requires an image to work on is skipped when there is no image attached. A model that requires at least one reference is skipped when the tray is empty. Every skip comes back with the reason, naming the model and what it wanted — an earlier version of this dropped models silently, and a five-tile run once produced twelve jobs from one model and none from the other with no message anywhere.
The Presets tab on the composer is worth a look before you build a comparison by hand. Each preset is a prompt, an aspect ratio, a copy count and a set of models chosen for the job — the product preset, for instance, fires Seedream 4.5, Nano Banana Pro and GPT Image 2 at a square frame with two copies each. Loading one and then swapping a model in or out is usually faster than starting from an empty composer, and the previews show you what the recipe actually produces.
There is no fixed ceiling on the selection. What limits you in practice is your plan's parallel job allowance and your wallets, both of which the quote on the button accounts for before you press.
It charges once per model, because it is running each model separately. That is the point of the comparison, and it is why the button shows a total and a wallet split rather than a single per-image price.
No. Hosted models bill their provider per call and green credits cannot pay for them, so they need gold — from a paid plan's monthly grant or a credit purchase. The own-GPU rows and GPT Image 2 are the free comparison.
Our own GPU runs one job at a time and hosted providers run yours immediately. In a mixed press the local rows are usually the last to land, and the card shows the real queue position rather than an indefinite spinner.
Yes. Regenerate on a finished card restores the full input set — prompt, references, ratio, model — so the follow-up run starts from exactly what made the picture you liked.