Would be great to see time spent per task per model. Especially since article references 390+ tokens per second for GLM5.2.