There are objective metrics, that’s what benchmarks are.
These companies aren’t just making stuff up, they really do want to improve the models, and the models really are improving.
I’m aware of multiple cases of benchmark cheating/“optimization”. So, taking benchmarks a face value seems laughable.