I think different domain benchmarks already exist and are widely used.

True, but afaik none that an AI company itself uses to say 'this is what we think showcases the individual strengths of our models the best'.