Reasoning benchmarks have converged across LLM vendors. What still differs, refusal rates, cost-to-control ratio, and jurisdiction, determines which model act