Qwen Image Max vs Qwen Image on Sume: 3.75x the price, no references

Qwen Image Max costs $0.09375 per image on Sume, which is 3.75 times the $0.025 of Qwen Image, and it takes no reference images, while Qwen Image takes up to 10. Both list 13 aspect ratios. So the Max row is for text-to-image jobs where you have compared the output and prefer it, not for edits.

The figures come from the catalog code on origin/main, read on 2026-10-10, and the Image API docs. The catalog does not say which row produces better pictures, so this post compares the contract, not the quality.

Side by side

Neither row lists resolution, quality, mask_url or background. Both list the same count of ratios. The two differences that matter are the price and the reference limit.

What the premium costs at volume

For 100 images the standard row costs $2.50 and the Max row $9.375. For 1,000 it is $25 against $93.75. The gap is $68.75 per 1,000 images, which is trivial for a hero set and large for a catalog run.

The reference limit matters more than the price. A product-photo workflow that starts from a supplied image cannot use the Max row at all, because a request with input_references returns a 400 on a row whose descriptor lists a maximum of 0.

A simple way to decide

Generate the same five prompts on both rows. If you cannot tell which one is the Max output when the labels are hidden, use the cheaper row. If the Max row wins on the cases you care about, use it for those prompts only, and keep the cheaper row for drafts and edits.

Keep a note of the date of the comparison. Catalog prices and the models behind a row can change, and the pricing lines are the source of truth for what Sume bills.

• Use Qwen Image for edits and drafts.
• Use Qwen Image Max only for text-to-image where you prefer it.
• Pin the row id; do not rely on sume/auto.
• Recheck prices with the endpoints route before a large run.

Checking the numbers yourself

Do not take the table on trust. Call the list route and read each row's supported_parameters, then call the endpoints route and read the first pricing line. The ratio of the two prices is 0.09375 divided by 0.025, which is 3.75. The reference limits come from the input_references descriptor, where the Max row shows a maximum of 0.

Repeat the check when you plan a big run. The catalog is code, and rows change as providers change.

Where text-only rows fit

A text-only row is a good fit when every image is described in words: stock-style art, backgrounds, concept sketches. It is a poor fit when the image must contain a product you already have. For that case the Qwen Image row, with its reference limit of 10, or one of the other 10-reference rows is the right starting point.

Budget example

A team that makes 300 text-to-image concept images a month spends $7.50 on Qwen Image and $28.125 on Qwen Image Max. The $20.625 difference is easy to absorb if the Max output saves an hour of rework, and hard to justify if no one can tell the two apart. Run the blind comparison once, write the decision down with the date, and revisit it when the catalog changes.

If you only need the Max row for a handful of hero images, call it by explicit id for those and leave the default on the cheaper row.

Whatever you choose, record the decision with the row id and the date in your repository, so the next person does not repeat the test blind and can see which prompts you used.