They used their own proprietary benchmark called sudo L7, which purports to evaluate staff-level engineering judgment more than the correctness of the code itself.

Not surprising to me that higher thinking will outperform coding harnesses here.