FLUX.2 pro or flex on Sume: $0.0375 vs $0.0625 per image

On Sume, black-forest-labs/flux-2-pro bills about $0.0375 per image and black-forest-labs/flux-2-flex about $0.0625. Both list 13 aspect ratios and take up to 10 reference images. Flex costs two-thirds more, so pay for it only if a side-by-side test on your own prompts shows a difference you can see. Black Forest Labs' models page lists the FLUX.2 family, with Max and Klein variants alongside, though only the pro and flex rows are what this post prices on Sume.

Do not choose by price alone, and do not choose by a launch headline. Run the same ten prompts on both rows and compare the images.

Side by side

A ten-prompt test

• Pick ten real prompts from your backlog, including two with text in the image and two with a reference photo.
• Send each to both rows at the same aspect_ratio; the whole test costs $1.00 for 10 pairs.
• Put the pairs in a grid with the row name hidden and ask two people which they prefer.
• If flex wins fewer than about six of ten, use pro. Set your own bar before you look.

Cost at your volume

The 2.5 cent gap looks small, but it scales. A product team that makes 2,000 images a month spends $75 on pro and $125 on flex, a $50 difference. For a campaign with 200 images, the gap is $5, and the choice should follow quality.

Both rows are billed from the catalog list times 1.25 per image, and a failed generation is not billed. Read usage.cost on each response and total it, so the monthly figure comes from your own logs and not from this table.

Request shape

The two rows differ by model id only. Swap the id and keep the rest:

{
  "model": "black-forest-labs/flux-2-flex",
  "prompt": "studio photo of a ceramic mug on a linen cloth, soft window light",
  "aspect_ratio": "4:5"
}

Reading the result

Compare the winners on the details that differ between the two rows: fine texture, hands, small text, and how closely each follows a long prompt. Keep the notes in a file with the prompt id, since a month later you will not remember which grid was which. If the rows tie, take the cheaper one and move on; a tie is also a result.

When the vendor ships a new FLUX version, rerun the same ten prompts against the new row instead of rewriting the test. The catalog at GET /v1/images/models shows which rows are live.

What Sume does not do

Sume does not expose provider-specific controls on these rows; provider.options must be empty, and the catalog is the allowlist, so a parameter that flex may support upstream is rejected here with unsupported_parameter if it is not in the catalog. Sume also does not carry the FLUX.2 Max or Klein variants discussed on the vendor page; check GET /v1/images/models for what is live.