This study analyse the potential of Large Language Models (LLMs) for social simulation by evaluating their ability to: (a) align with principles of economic rationality; (b) make decisions consistent with explicit preferences; (c) exhibit socially rational behaviour; and (d) refine beliefs to anticipate the actions of other agents. Using a series of game-theoretic experiments, this paper assess the performance of models such as GPT-4.5, Mistral-Small, and Qwen3. The results indicate that while some models behave consistently in simple settings, they struggle in more complex scenarios that require strategic anticipation. This study highlights both the promise and the current limitations of LLM-based generative agents in computational social science.