73 SKUs through generate, cutout, upscale: upscale is 75-81% of cost

A product-shot pipeline of generate, remove background and upscale costs between $0.24725 and $0.26625 per SKU on Sume, depending on the image model. For 73 SKUs that is $18.05 to $19.44. In every variant the $0.20 upscale is the largest line, at 75 to 81 percent of the SKU cost.

Per-SKU and 73-SKU totals

Each SKU pays three fixed lines: one generation at the model's price, one cutout at $0.0225 and one upscale at $0.20. Composition in Pillow or a design tool is free. The generation prices come from the shared Sume list: Grok Image $0.025, GPT Image 2.5 low $0.02475, and Seedream 5 Lite $0.04375.

What this means for the generator choice

Choosing between these three models moves 73 SKUs by $1.39 at most ($19.43625 minus $18.04925). Skipping the upscale on 50 of the 73 SKUs moves it by 50 x $0.20 = $10.00. The upscale decision matters about seven times more than the generator decision.

Without any upscale, the three pipelines cost $3.4675, $3.44925 and $4.83625 for 73 SKUs. If the listing page shows images at 1000 pixels, that may be the whole job.

Where to put the upscale

Upscale only what a customer will see large. A zoomable gallery image deserves it. A search-result tile does not. Decide by placement, tag each SKU row with a needs_upscale flag, and submit upscales from that column.

If you do upscale a transparent cutout, read the order guide first, because it explains how to keep the alpha channel through the step. The Python walk-through of generate then cutout has the two-call code.

A tiered plan

Put the 73 SKUs in three tiers. The top 8 get the full pipeline. The next 25 get generate and cutout only. The remaining 40 get generate only. With GPT Image 2.5 low, that is 8 x $0.24725 + 25 x $0.04725 + 40 x $0.02475 = $1.978 + $1.18125 + $0.99 = $4.14925.

That is under a quarter of the $18.04925 full-pipeline cost, and the pages that earn the most revenue keep the best image.

Checking the numbers yourself

Every figure above is a product of two catalog prices and a count, so you can reproduce it in a spreadsheet. Pull the prices from the public catalog on the day you run the batch, because list prices change. Then multiply by the number of SKUs in each tier.

Keep one row per SKU with columns for generator, whether it was cut out, whether it was upscaled, and the three job ids. After the run, sum the usage.cost values from the responses and compare the total with your estimate. A gap larger than a few percent means a retry loop or a wrong tier assignment, and it is much easier to find in a 73-row sheet than in an invoice.