In this study, we quantified how effectively four large language models (LLMs)—GPT-4o, GPT-4-preview-1106, GPT-4o mini, and GPT-3.5 Turbo—mimic human social behaviors using the "Negotiation Arena" framework. Drawing from recent research that highlights the need for complex conversational task simulations to evaluate LLM performance, we utilized the "Buy and Sell" game to assign distinct personas (Cooperative, Competitive, Altruistic, Selfish, Cunning, Desperate, Control) to each model. Our findings show that end-to-end multimodal models like GPT-4o and GPT-4o mini exhibit strong persona-driven behavior, while text-based models like GPT-4-preview-1106 also respond to persona instructions, though with different tendencies. Aggressive and self-centered personas performed better in negotiations, while altruistic and cooperative personas showed lower success rates. This study provides insights into how LLMs can replicate human-like persona behaviors, emphasizing the potential of persona-driven simulations as an evaluation method for practical applications in social interactions.