Text-in-image bake-off: 20 prompts on 4 Sume models costs $7.00

Twenty text-heavy prompts on four Sume models cost $7.00: $1.60 on Ideogram 4.5 at medium, $1.40 on GPT Image 2.5 at high, $2.00 on Qwen Image Max and $2.00 on Nano Banana 2.1. Sume does not publish a text-rendering ranking, so a small bake-off on your own words is the honest way to choose.

The bake-off budget

Each model generates the same 20 prompts at one image per prompt. Sume bills the catalog list price multiplied by 1.25 and rounded up to the next cent.

How to run it so the result means something

Use words that matter to you: a real headline, a price, a product name, a date. Include a short word, a long word, numbers and one accented or mixed-case string. Keep the ratio and the prompt identical across models, and generate each prompt once per model so the comparison is not luck.

Score by reading, not by eye. Copy the text out of each image or check it by hand against the prompt, and count exact matches. If a model gets 18 of 20 right at medium, there is little reason to buy a higher tier.

• Ideogram 4.5 has low, medium and high tiers at 4, 8 and 28 cents billed, so re-run the misses at high.
• GPT Image 2.5 can be raised to xhigh or max for the misses.
• Qwen Image Max cannot be fixed with an edit, since it is text-to-image only.

After you pick

Put the rare, exact text in code when the stakes are high, such as a price or a legal line, and let the model make the artwork. For graphics where the model's text is good enough, keep the winner and move to the 100-image run.

Reading the cost correctly

The $7.00 total assumes one image per prompt per model. Doubling to two takes per prompt gives $14.00 and lets you see variance. Re-running only the failures at a higher tier is much cheaper than starting all over at that tier: if Ideogram 4.5 misses 4 of 20 at medium, the retries at high cost 4 x 28 = 112 cents, or $1.12.

Record the model id, quality, ratio and prompt next to each result in a sheet. Without that record a month-old winner cannot be reproduced, and Sume does not serve seed on image models in v1, so a repeat run will not return the same picture. Keep the winning prompt, not the winning file, as your asset.