FLUX.2 pro or flex for label text on a bottle: Sume ids

Black Forest Labs describes FLUX.2 [flex] as its typography and detail specialist and [pro] as the affordable production-grade option. On Sume the pro id is black-forest-labs/flux.2-pro; confirm the flex id appears in GET /v1/images/models for your key before you build on it. For a bottle label with real words on it, try flex first and compare against pro on the same prompt, then decide on cost.

What the vendor page says

The FLUX.2 [max] page positions max as the maximum-performance tier and names the alternatives: pro, flex and klein for fast prototyping. It says max has the highest editing consistency and suits product photography. The page did not list prices or exact reference counts in the part I read, so this post gives none for BFL.

A fair A/B on label text

Use one prompt with the exact label string in quotes, the same aspect ratio and the same reference photo. Run each model three times, because a single sample hides variance. Sume's Image API has no seed support in v1, so you cannot lock randomness; repeat runs instead.

for m in black-forest-labs/flux.2-pro black-forest-labs/flux.2-flex; do
  curl -s -X POST "https://api.sume.com/v1/images" \
    -H "Authorization: Bearer $SUME_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"model\":\"$m\",\"prompt\":\"amber glass bottle on a white sweep, front label reads 'COLD BREW 250 ML' in black sans-serif\",\"aspect_ratio\":\"3:4\",\"n\":3}"
  echo
done

Reading the result

Count the images where every character is right, not the ones that look nicest. Then read each model's pricing line from its endpoint record and divide the price by the number of correct images. The cheaper model per correct image wins, which may not be the cheaper model per call.

Limits

Check each model's n range descriptor before asking for three images in one call, since per-model ceilings are lower than the 10-image request cap. If neither FLUX tier gets the label right, try another model from the image models list and score it the same way.

Test setup

Keep the comparison fair:

• Same prompt, same ratio, same reference photo.
• Three runs per model.
• Score each image pass or fail on the label string, not on looks.