In the low-cost range of approximately $0.25 per task, the Chinese and American models performed similarly. As the cost rose to approximately $1, the American model led in intelligence by about 15 points, and remained ahead by about 12 points in the high-end range. The performance frontier of the Chinese model plateaued around 46 points, and after the cost per task exceeded approximately $3.5, further increases in expenditure did not bring a significant score improvement; the American model, on the other hand, could achieve around 51 points with a cost of approximately $1. The results show that the Chinese model remains competitive in low-cost applications, but in the high-capability range covered by this chart, the American model is better able to convert inference expenditure into performance gains. However, for enterprise users, simply comparing the price per million tokens is far from sufficient, as the actual cost and capability of different models to complete the same set of tasks vary considerably.
