[GLM 5.2 Reproduces Academic Papers at Approximately One-Eighth the Cost of Opus 4.8] According to monitoring by Beating, the open-source large model GLM 5.2 demonstrated exceptional cost-effectiveness in academic reproduction tests. The alphaXiv team used automated agents to test the ability of large models to reproduce cutting-edge papers. When reproducing the self-distillation reinforcement learning paper SDPO, the operational cost of GLM 5.2 was only about one-eighth that of the closed-source flagship model Claude Opus 4.8 Max. The experiment required the model to autonomously read the paper, troubleshoot complex environment errors in the open-source library VeRL, and complete ablation experiments. GLM 5.2 succeeded in reproduction after 14 failed runs, consuming 2.65 million tokens at a total cost of $6.21. Claude Opus 4.8 Max succeeded after 9 failed runs, consuming 4.53 million tokens at a cost of $46.35. [Original Link]