From Benchmarks to Real-World Performance: A Data-Driven Assessment of Large Language Models in 2025