One Agent Benchmark Puts Nvidia 5x Ahead Of AMD On Cost

SemiAnalysis released AgentX, an open-source benchmark replaying real coding-agent sessions to evaluate inference stack performance. Using GLM 5.3 on Nvidia hardware, it found Nvidia up to 5x more cost-efficient than AMD at 150 tokens/sec per user, highlighting a significant cost gap in multi-turn agentic inference workloads.