“Large language models now achieve over 90% accuracy on popular benchmarks like MMLU.”