Rippling's real-payroll AI test: cheap GLM 5.2 ties premium on accuracy at half the cost
Rippling's President & CPO published a real production test of 15 AI models doing scored payroll agent work — 2,100 graded runs each, with accuracy and cost recorded. The cheapest model, GLM 5.2, tied the most expensive on accuracy at half the price, while quality flattened at the top and cost diverged sharply.