I think it's because the tests are more for general intelligence, but 5.1 is likely aimed more towards agentic coding and tool usage, so it lost a bit of intelligence in exchange for consistency and more focus coding use-case.