Open-source models are crushing it in cost-effectiveness, doing the same job at one-eighth the cost—research budgets no longer have to be all-in on closed-source.

BiJieNet news: OneMillion_AI reports that the open-source large model GLM 5.2 demonstrated extremely high cost-effectiveness in academic reproduction tests. The cost of reproducing its associated academic paper is only about one-eighth of the closed-source flagship model Claude Opus 4.8 Max. The research platform 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, GLM 5.2 succeeded after 14 failed runs, consuming 2.65 million tokens and costing $6.21, while Claude Opus 4.8 Max succeeded after 9 failures, consuming 4.53 million tokens and costing $46.35.