Rehearse an image prompt for 2.5 cents, then render the keeper

Yes: try the wording of a prompt on a model that costs 2.5 cents an image, then send the winning prompt to the expensive model once. At Sume's catalog base prices, eight drafts on Qwen Image plus one final on Nano Banana Pro cost 0.3875 USD, while nine finals on Nano Banana Pro cost 1.6875 USD.

The saving is real only if the cheap model reads your prompt the way the expensive one does, so the method needs one honest check at the end: the final render still has to be judged on its own.

The price ladder

The catalog base prices below were read on 2026-10-10. A request can cost more than the base price when you raise a quality or resolution field, so use the endpoint pricing lines for the exact settings you plan to send.

The arithmetic

Eight drafts at 0.025 USD is 0.20 USD. Add one final at 0.1875 USD and the total is 0.3875 USD. Nine finals at 0.1875 USD is 1.6875 USD. The difference is 1.30 USD, or about 77 percent less for the same nine attempts, as long as the ninth attempt is the one you keep.

Failed and cancelled generations are not billed, so a draft that errors out costs nothing. A draft that succeeds but looks wrong costs its full 0.025 USD, which is still cheap.

Which draft model to use

Pick a draft model that matches the final model's inputs. If the final render needs a reference photo, the draft model must accept references too, and Imagen 4 Fast does not: it is text-to-image only, and a request with input_references returns 400 unsupported_parameter. Qwen Image and Grok Image both take references.

Match the ratio as well. Nano Banana Pro lists 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3 and 9:16 plus auto. Qwen Image lists most of these but not auto, so a draft for an edit should use a fixed ratio from the final model's list that Qwen also lists, such as 4:5.

Use n: 4 on Qwen Image to see four takes per call; Grok Image allows only one image per call.

Rules for the draft step

Keep the draft model fixed for the whole project, so that you compare wordings and not models. Save every prompt and the take you chose, and send the prompt text unchanged to the final model. If you edit the prompt between the draft and the final, you have thrown away the thing you paid to test.

Also keep the ratio and the references identical. A draft at 1:1 tells you little about a final at 9:16, because composition changes with the frame. If the final needs aspect_ratio: "9:16", draft at 9:16, which Qwen Image and Grok Image both list.

What the cheap draft can and cannot tell you

A draft is good at showing whether the idea works: composition, subject, setting, mood, what to leave out. It cannot tell you how the final model will render fine text, skin, or small product detail. Treat anything in that group as unverified until the final render.

A simple routine keeps the saving honest:

• Write the prompt in plain terms and run four drafts with n: 4.
• Change one thing, run four more, and stop after two rounds.
• Send the best prompt, with the same references and ratio, to Nano Banana Pro once.
• If the final disappoints, change the wording in the draft model first, not in the expensive one.