A lot of it comes down to the context and a prompting strategy tailored to the particular model. I don’t believe the current benchmarks really take those optimizations into account.

I’ve personally been getting better results with Gemini as well, but I think it’s just because I’ve used it more.