One area where I anecdotally see significant improvement with o3 is image recognition. I used the image you see here in class on Tuesday (pre o3). All of the models had issues but overall Gemini 2.5 did best.

Retested using o3 and it nailed it. Every title, author, and edition was perfect.