Sale banner with exact price digits: low drafts, one high, 11 cents
Test a sale banner's price digits with four low-quality GPT Image 2.5 drafts (4 cents), then render the winner once at high (7 cents). Total 11 cents on Sume.
For a banner with exact price digits, generate four drafts at quality: "low" (4 cents), read the digits on each, and render the best prompt once at high (7 cents). The total is 11 cents on Sume, and you check every digit by eye before anything is published. OpenAI's image guide says that text rendering has improved but the model can still struggle with precise text placement and clarity.
The two-step cost
Both steps use ChatGPT Image 2.5 (openai/gpt-image-2.5). The figures are the catalog's rounded-up cents.
| Step | Call | Cost |
|---|---|---|
| 1 | n=4, quality low | 4 x 1 cent = 4 cents |
| 2 | n=1, quality high, same prompt | 7 cents |
| Total | - | 11 cents |
How to write the prompt
Put the number in quotes and say it in words as well: '$19.99, nineteen ninety-nine'. Give the text one location and one size ('large, top left') and keep the line short. Avoid long strings of small print inside the picture; put legal copy in your layout tool.
Specify the banner shape with a ratio the row lists. GPT Image 2.5 lists 16:9, 4:3, 1:1, 4:5 and 9:16, but not 21:9 or 3:2. For a very wide web banner, render at 16:9 and extend the canvas in your design tool.
- Quote every string:
"SALE","$19.99","ENDS SUNDAY". - Name the font style in plain words: bold sans-serif, white on dark.
- Ask for no other text.
- Check every digit and every letter against your source sheet.
The digit check
This post makes no claim about how often low-quality drafts get digits right; measure it on your own prompt. Count correct digits across the four drafts; if none is right, tighten the prompt and run another four for 4 cents rather than spending 7 cents on a high render that may repeat the error.
When the high render comes back, compare it to your source sheet again. Higher quality does not guarantee the digits survive.
If digits keep failing
Use the render as a background and add the numbers in a design tool or with an image library. Price digits are a legal matter on many storefronts, and a single wrong digit costs more than any number of tests. Ideogram 4.5 is the other row worth trying for text: its medium tier is 8 cents, and a pair of low drafts is 8 cents.
Keep a proof sheet
Write the exact strings in a text file next to the generated image and tick them off. Two people checking is better than one for numbers. Archive the prompt, the model id and the usage.cost with the approved image, so a later question about a price on a banner can be answered from the record.
Scaling the plan to a campaign
For a campaign with 12 price points, the draft and final steps cost 12 x 11 cents = $1.32 if every prompt succeeds first time. Allow a second draft round for a third of them, 4 x 4 cents = 16 cents more. The whole campaign's image cost is still under $1.50, which is why the digit check, not the generation, is where your time should go.
Sources
Related posts
More in Use cases
- School open-day teaser: three 10 s MiniMax H3 Max takes cost $3.00
Three 10-second H3 Max takes at 768p are $1.00 each on Sume, with reference photos of your building. H3 and 1080p comparisons and the limits.
- Seedance draft to final: Mini 480p drafts, 2.5 1080p finals, priced
Draft 20 shots on Seedance 2 Mini at 480p, then re-render the approved few on 2.5 at 1080p. Total cost for 3, 6 and 10 finals, with the request change.
- 50 four-second hook variants on Seedance: $17.577 to $115.56
Fifty 4-second hook tests on Sume cost from the Mini 480p price to the 2.5 720p price. Per-tier totals at three resolutions and a submit script.
- Seedream 5 Lite from a reference photo: 4.4 cents per product image
seedream-5-lite costs $0.04375 per image on Sume with up to 10 references and n up to 4. Hero-to-variants request, plus a 100-product bill next to Seedream 4.5.
Written by Sume