These kinds of benchmarks are misleading without a joint metric showing much work was necessary by humans after the fact. How much time to clean up that 2h42m of code? Style and architecture need to make sense, not just passing tests. That's the bottleneck now: reviewing!