This model is horrible right now. The newly released 3.5 Flash High has incredibly low performance.
When I asked it to check back, I explicitly told it NOT to delete anything. Instead, it went ahead and deleted all my code.
The absolute worst part is that when I clicked confirm undo, it didn’t even show that the code had been deleted!
I then asked it to check exactly what code it had just deleted. It had completely no idea, and just kept continuously running other files.
With this kind of performance, you are thinking about beating GPT 5.5? It’s not even as good as 4.0!