Recraft V4 or ChatGPT Image 2.5 for marketing graphics on Sume

Use Recraft V4 on Sume for a quick, fixed-price graphic from a text prompt, and ChatGPT Image 2.5 when the graphic needs a transparent background, a mask edit or reference photos. Recraft V4 bills $0.05 per image, is text-to-image only and returns webp only. ChatGPT Image 2.5 is token-billed, accepts up to 16 references, an optional mask_url, and background: transparent.

fal's Recraft V4 page (read 2026-10-08) lists $0.04 per image and describes the model as developed with designers, licensed for commercial use and aimed at production workflows. It does not say whether vector or SVG output exists on that endpoint, and Sume's row lists webp only.

Side by side

Everything in the Sume columns is from the catalog and docs on origin/main.

The webp point

Because Recraft V4 returns only webp, a pipeline that expects PNG needs a conversion step. A request that sets output_format to a value the row does not list is rejected with 400 unsupported_parameter; the catalog lists webp and nothing else. Check that your design tool, store listing or ad platform accepts webp before you build on this row.

How to choose

Price per attempt matters less than attempts per keeper. A fixed $0.05 is easy to budget for a set of illustrated tiles, while a token-billed row moves with size and quality, so run a small test and total usage.cost. If the finished graphic must sit on a coloured page, ask ChatGPT Image 2.5 for a transparent background and check the alpha channel before shipping.

For a set that must stay consistent, generate one anchor image first and pass it as a reference to ChatGPT Image 2.5. Recraft V4 cannot take one on Sume, so the consistency has to come from the prompt.

Before committing a design system to either row, run the same five briefs on both and keep the files. Check the results at the size they will be shown, not zoomed in, and check that the background, edge quality and any lettering survive the export format you need. The Sume catalog tells you what each row accepts; it does not tell you which output suits your brand.