Also doesn't the very good ARC AGI 2 score of GPT-6 Astra kinda contradict this, since each problem is its own game with very different rules