I understand what you’re saying, but I am generally very careful when comparing models and their benchmarks; benchmarks often don’t really match “real world” quality.