GPT Image 2.5 misspells a brand name: spell it letter by letter

If GPT Image 2.5 misspells a brand name inside an image, put the exact wording in quotes, spell the tricky word out letter by letter in the prompt, and ask for no other text. Then read the result yourself, because OpenAI's own docs say the model can still struggle with text placement and clarity. On Sume the same prompt goes to openai/gpt-image-2.5 through POST /v1/images.

What OpenAI's docs say about text in images

The table below is limited to what OpenAI's current pages say. None of it is a guarantee of accuracy, and OpenAI's image guide still lists text placement and clarity as a limitation.

Source pages: Image prompting, the prompting guide in the OpenAI Cookbook and the Image generation guide.

A prompt that follows that advice

Name the surface, name the text once, give the spelling, and close the door on extra words. The invented word below is only there to show the pattern; swap in your own.

The request below keeps quality at high. On Sume, an omitted quality already defaults to high for this model, so setting it is a way to make the intent visible in your code.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Poster, flat navy background. Centered headline, exactly once: \"KOHLRABI\" (K-O-H-L-R-A-B-I), white bold sans-serif. No other text.",
    "quality": "high",
    "image_size": "1024x1536"
  }'

Check the output before it ships

Sume's image response is a hosted image URL plus a usage block. It has no field that reports the text found in the picture, so the spelling check is a human or a tool of your own. Open each result at the size it will be shown at, and compare the letters against your source string.

If a word is wrong, change one thing at a time: first the letter-by-letter spelling, then the quality tier, then the placement wording. OpenAI's iteration advice is to refine with small, single-change follow-ups and compare results before adding more instructions.

• Is every letter correct, including accents and capital letters?
• Does the text appear only where you asked, and only as many times as you asked? See why a headline appears twice.
• Is it legible at the real display size, not only zoomed in?

When to switch models

A failed spelling after two or three careful prompts is a sign to compare another model, not to keep retrying the same one. Failed Sume image generations are not billed, but a completed image with the wrong spelling is billed in full, so each retry on a finished image has a cost.

Sume's Image API docs say each model advertises the parameters it accepts, so read the catalog before you pin a model. For a side-by-side of catalog models that handle text, see the text-in-images comparison.