The problem with Gemini 3.0 is that it doesn’t perfectly follow system instructions. While Gemini 2.5 adhered to them perfectly, and the benchmarks look impressive, I think failing to follow instructions is a real issue. I know it’s still early days and updates are ongoing, but I hope they fix this soon to match 2.5’s adherence.
They have documented how to approach Gemini 3.0. It is different from 2.5.
You need short, clear prompts, not detailed instructions, less “step-by-step” guidance. Also, you need to keep the temperature at 1.0, otherwise it might break reasoning and make performance worse. I am yet to test it today.
What I am disappointed is that they made the AI Studio paid now and Gemini 3 is quite expensive especially if you use documents like PDF. It takes literally 2-3 times more tokens. I ran 2 prompts yesterday for 2$+. That’s an insane costs…