Gemini seems much better than GPT-4 in image understanding.
Here's a test I did trying to read Persian from an image:
https://chat.openai.com/c/bc45cc9c-fd19-4359-9096-3936d4d17a...
https://gemini.google.com/app/3286bdfb5f24cd36
I can read Persian and can confirm Gemini (not even the Advanced model) got it right while GPT-4 blew it.