If they "knew the questions in advance," why'd they need Internet access at all? The ability to use the same data sources humans would use is not the insult you seem to think it is.
Again: the assertion was yours, so let us know the results of your own work.
Without internet: 10%
With internet: 23%
In addition:
> We found that the ground-truth answers for one dataset were widely leaked online
in very small letters, and they blocked these URLs at runtime but not training time.
It's not bad, but not revolutionary at all compared to the leap that was GPT-2 from GPT-3, or GPT-4o to DeepSeek-R1