This paper introduces a benchmark and a buyer-enhancement method for evaluating the bargaining abilities of large language models (LLMs). The Bargaining task is formally defined as an asymmetric incomplete information game, where the Buyer and Seller have different goals and private information. A real product price dataset, AmazonHistoryPrice, is collected to evaluate various LLMs' bargaining abilities. The results show that playing as a Buyer is significantly more challenging than as a Seller, and increasing model size does not effectively improve the Buyer's performance. To address this challenge, a novel approach called OG-Narrator is proposed, which integrates a deterministic Offer Generator to control the price range of Buyer's offers and an LLM Narrator to generate natural language sentences for generated offers. Experimental results show that OG-Narrator significantly improves the buyer's deal rates and profits. The paper also defines metrics to evaluate the bargaining abilities of LLMs, including Normalized Profits (NP), Sum of Normalized Profits (SNP), and Share. The results indicate that larger models perform better as Sellers but not as Buyers. The OG-Narrator method enhances the Buyer's performance, even for unaligned models, and demonstrates that the ChatGPT Seller is vulnerable when facing a Buyer enhanced by OG-Narrator. The paper also discusses the limitations of the dataset and the challenges in evaluating bargaining abilities. The findings suggest that the Buyer's role is more complex and challenging than the Seller's, and that the OG-Narrator method is an effective way to improve the Buyer's performance in bargaining tasks.This paper introduces a benchmark and a buyer-enhancement method for evaluating the bargaining abilities of large language models (LLMs). The Bargaining task is formally defined as an asymmetric incomplete information game, where the Buyer and Seller have different goals and private information. A real product price dataset, AmazonHistoryPrice, is collected to evaluate various LLMs' bargaining abilities. The results show that playing as a Buyer is significantly more challenging than as a Seller, and increasing model size does not effectively improve the Buyer's performance. To address this challenge, a novel approach called OG-Narrator is proposed, which integrates a deterministic Offer Generator to control the price range of Buyer's offers and an LLM Narrator to generate natural language sentences for generated offers. Experimental results show that OG-Narrator significantly improves the buyer's deal rates and profits. The paper also defines metrics to evaluate the bargaining abilities of LLMs, including Normalized Profits (NP), Sum of Normalized Profits (SNP), and Share. The results indicate that larger models perform better as Sellers but not as Buyers. The OG-Narrator method enhances the Buyer's performance, even for unaligned models, and demonstrates that the ChatGPT Seller is vulnerable when facing a Buyer enhanced by OG-Narrator. The paper also discusses the limitations of the dataset and the challenges in evaluating bargaining abilities. The findings suggest that the Buyer's role is more complex and challenging than the Seller's, and that the OG-Narrator method is an effective way to improve the Buyer's performance in bargaining tasks.