We employ machine learning (ML), deep learning (DL), and fine-tuned large language models (LLMs) to construct climate variables from Corporate Social Responsibility (CSR) disclosures, ESG reports, and climate risk alerts of A-share listed firms in China. Our results show that LLMs consistently outperform traditional ML and DL approaches across all tasks, achieving the highest accuracy and F1 scores. Support Vector Machines (SVM) also deliver competitive performance in simpler classification settings, indicating their potential as a cost-effective alternative in resource-constrained environments. We further examine the relationship between LLM-generated climate variables and stock returns, revealing that while climate opportunity, climate and transition risk are priced by markets, overall climate-related content and physical risk are not. These findings suggest that China’s equity market selectively prices climate information.