$5 test wallet: one n=10 call at 4K can cost $2.00 on Nano Banana

The short answer

On Sume, a Nano Banana 2.1 call at 4K with n set to 10 costs up to $2.00 (10 x $0.20), so a $5 test wallet covers two such calls and a half. The same call at 0.5K costs $0.75, and a GPT Image 2.5 low call with n at 10 costs $0.2475.

Image Router has a wide price band, and a loop that sends the highest settings can empty a small wallet quickly. Work the worst case per call before the loop starts.

The arithmetic

Sume bills cost_usd times n, so the worst case per call is the largest price your request could select times the n you send. The table fixes n at 10 and varies the model and tier, then divides the $5 wallet by the call cost.

Divide $5.00 by each call cost and round down, because a call you cannot afford is not worth starting.

How to run it

Start with n set to 1 and a 0.5K or low draft, check the cost field in the response, and only then raise n. The usage block reports the billed amount in dollars.

Remember that auto quality on GPT Image 2.5 reserves max, so a wallet check that uses the auto reserve is higher than the high-tier price in the table.

Retries and what to watch

A wallet cap is only useful if the script checks it. Before each call, multiply the price of the tier you sent by n and compare it with the balance you have left, and stop when the next call would not fit.

Two habits keep the $5 from going on one bad loop.

• Send n set to 1 for the first call of every run and read usage.cost in the response.
• Cap the number of calls in the loop to the figure in the last column of the table.
• Do not use sume/auto in a capped loop, because the price band is wide and the chosen model is not disclosed.

What the docs say to check

On GPT Image 2.5 the quality values are auto, low, medium, high, xhigh and max. If you omit quality the default is high, and auto reserves max, so pin the tier you budgeted for instead of leaving it to the default.

The n parameter takes up to 10 images per call, but per-model ceilings are lower, so read the n range descriptor for the model before you plan the number of calls.

Keep one fixed model id in the loop. With sume/auto the model is not disclosed, so you cannot bound the per-image price by choosing a tier.

Every price in this post is a snapshot dated 2026-10-08. The Sume figures come from the public catalog and the Image API docs, the Google figures come from its Gemini API pricing page, and no figure here is a benchmark of quality or speed. Prices and model lists change, so confirm them on the day you spend money.