LLM Benchmark leaderboard for common evals sounds like a fun idea to me.

I haven’t come across any research showing that a specific LLM consistently outperforms others for this. It generally works best with strong reasoning models that produce consistent outputs.