how we test text in AI images:
1. hide a "wrong" detail in the sign, like "sourDOUGH".
2. same prompt, same 1K tier, one shot each. no rerolls.
3. read it letter by letter.
most models quietly fix the typo. the good ones copy it.
what would you hide in yours?
gpt image 2.6 vs 2.5, one casing trap.
the window had to read "CroiSSants 3.50 / sourDOUGH 6 / 1/2 off after 7pm".
2.6 (1) nailed every letter. it even turned 1/2 into a real fraction.
2.5 (2) wrote "SourDOUGH". so close.
6 credits vs 2 per 1K image. Which one do you prefer?
1
1
6