Done similar comparison recently with self hosted Kimi K2.6, glm 5.1 and DS V4 Pro (and small qwens as a bonus), max ctx, with minor technical advisory from my side. Kimi was the best one in terms of gtd, with a steady delivery through iterations, was also able to step back from the problem and solve it on a higher level (vs glm and ds). By default it was really over-thinking but the results (code) was actually less tangled than the one produced by the rest.