I think (2) is the important one. Anthropic provides benchmarks with their model releases. If they are making changes that make models perform worse on advertised benchmarks, how can this be legal?