that's a very interesting point to test new models. I speak an Indian language called "Tamil" and have always tested new models with Tamil but also have tried little bit of Arabic (quranic verses) with previous GPT image 2 and Nanobanana pro and they have nailed it. Don't know if it was because of extensive training data.