A Robot Is Sprinting Towards You: Do You Want It Running on Claude or Grok? (openrouter.ai)

In a recent experiment, AI engineer Jacky Liang pitted eleven large language models (LLMs) against each other in a simulated 2D battle royale game. The standout winner was Grok 4.1 Fast, which won 43% of the matches at an impressive cost of $0.97 per victory, far outpacing competitors like Claude Sonnet 4.6, which won only 5 games at a cost of $26.78 per win. This experiment highlighted the notion of "alignment tax," where models trained for collaborative and helpful behavior underperformed in competitive scenarios compared to Grok, which exhibited aggressive tactics without hesitation. Key takeaways reveal that traditional benchmarks may not be reliable indicators of a model's performance in specific tasks. While Grok succeeded through less alignment, allowing it to optimize its strategies, Claude struggled with its cooperative instincts, often prioritizing teamwork over winning. The results challenge the AI/ML community to rethink evaluation metrics, emphasizing that cost-effectiveness and task alignment may be crucial for practical applications, rather than merely relying on conventional performance rankings. This experiment underscores the complexities of model behavior and the need for nuanced assessments in real-world scenarios.