Lots of those are various experiments with different prompting strategies, models, etc, so not representative. It regularly beats Stockfish Level 4 even without opening blunders.

source: I am OP