Pick an image model by job: product photo, poster text, edit

Choose by the job. For a masked or transparent-background edit, use openai/gpt-image-2.5, the only model on Sume that takes mask_url and background. For text inside the image, try ideogram/ideogram-v4.5. For multi-reference product shots, try google/nano-banana-2.1 or black-forest-labs/flux.2-pro. For a hosted Qwen, use qwen/qwen-image-max. If you do not care, send sume/auto.

What shipped in the last five weeks

ChatGPT Images 2.5 and its two API models arrived on 2026-09-08, Qwen-Image 2.1 weights in September, Ideogram 4.5 on 2026-09-30, and Nano Banana 2.1 on 2026-10-06. Sume lists the GPT, Ideogram, Nano Banana, FLUX.2 and Qwen families, but not every vendor release is a Sume id, so the table points at the ids that exist.

The limits that decide it

Reference counts differ: 16 on GPT Image 2.5, up to 10 by default, and 5 on Ideogram 4.5. Quality tiers differ: xhigh and max are GPT Image 2.5 only. Parameters a model does not list return 400 unsupported_parameter, so a wrong pick fails loudly instead of quietly dropping your mask.

Test before you commit

Run the same prompt and reference on two candidates, three samples each, and compare the cost per usable image, using usage.cost from each response. Sume has no seed in v1, so repeated runs are the only honest test.

for m in openai/gpt-image-2.5 ideogram/ideogram-v4.5 google/nano-banana-2.1; 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\":\"studio photo of a matte black water bottle on a white sweep, label reads HYDRA 750 ML\",\"aspect_ratio\":\"1:1\"}" | jq -r '.model, .usage.cost, .data[0].url'
done

What this table does not claim

Vendor quality claims are vendor claims. Nothing here ranks the models on image quality, because I ran no benchmark. The prices quoted are list prices from the pages I read; Sume's own price per model comes from its endpoint record.

When to re-run this table

Update your choice when:

• A model is retired and its id starts running as a successor.
• A catalog row adds or drops a parameter you rely on.
• A price line changes in the endpoint record.