Advances in generative AI (GenAI) facilitated the development of AI-based non-player characters (NPCs) capable of interacting with players more dynamically. These NPCs further allow for natural language communication with players via chat or speech. However, little is known about how players interact with GenAI-based NPCs in collaborative tasks, particularly when NPCs exhibit different levels of autonomy. We developed a video game where players collaborate with an NPC across three levels varying in NPC autonomy: NPC strictly follows player instructions, NPC has autonomy but prioritizes player guidance, and NPC operates fully autonomously. We conducted a user study ( \(n~=~16\) ) to examine how players perceive and interact with NPCs under these conditions. Results show that while greater autonomy enhances helpfulness and collaboration, players prefer a balance between proactive AI support and personal control. Our findings provide insights into player preferences, engagement, and the perceived agency of AI-driven NPCs in collaborative settings.