Poster text test: Nano Banana Pro at 19 cents vs GPT Image 2.5 at 7

To choose between Nano Banana Pro and GPT Image 2.5 for a poster with text, run the same prompt on both and compare the spelling yourself: on Sume that costs 26 cents (19 cents for Pro at 1K, 7 cents for GPT Image 2.5 at high), or 52 cents for two prompts each. No vendor page settles this for your copy, and OpenAI's own guide says text rendering has improved but can still struggle with precise placement and clarity.

What each costs

Both models are pass-through rows in the Sume image catalog. The figures below are list prices from the catalog times 1.25, rounded up to the cent, for 1K or 1024-class output.

The test plan

Write three copy blocks of different difficulty: a short headline (three words), a headline plus a date line, and a headline plus five lines of small print. Run each on both models at the same aspect_ratio (both list 4:5 and 3:4). That is six calls and 78 cents in total (3 x 19 + 3 x 7 = 57 + 21 = 78).

• Send the exact text in quotes and say where it goes.
• Spell unusual words letter by letter in the prompt.
• Judge the result at the size you will publish, not zoomed in.
• Count wrong or missing characters per image and write the number down.

What differs in the request

Both accept input_references for edits. GPT Image 2.5 takes up to 16 references and an optional mask_url, and quality from auto to max. Nano Banana Pro takes resolution tiers 0.5K to 4K and a longer ratio list, including 21:9. Pro's 4K list is higher than its default, so check usage.cost when you ask for 4K.

If the poster needs a transparent background, only the GPT Image 2.5 request supports background: "transparent" among these two.

Cost of getting it wrong

A poster with a misspelled headline is not a 7-cent or 19-cent problem; it is a reprint or a reshoot. That is the argument for spending the 78 cents up front. A single misspelling found after a print run costs more than every image test you could run in a month.

Keep the test prompts in your repository, run them again whenever the catalog adds a row or a model version changes, and compare against the saved scores. The catalog changes: Sume has already moved Nano Banana 2 requests to Nano Banana 2.1, and GPT Image 2 stays selectable beside 2.5.

Reading the outcome

If both models spell the short headline right and diverge on the small print, small print belongs in your layout tool, not in the image. If GPT Image 2.5 gets all three right at 7 cents, there is no case for paying 19 cents for the same job. If Pro wins, you now have a measured reason to use it for the hero and a cheaper row for drafts.

Because sume/auto hides the family it ran, pin both ids for this test so the comparison is clean.