Nvidia shows high cost efficiency in new AI agent benchmark

A new open-source benchmark, AgentX, released by SemiAnalysis, indicates that Nvidia hardware may offer significantly higher cost efficiency for AI agentic sessions compared to AMD. The benchmark, which uses real coding-agent sessions rather than fixed-length prompts, reports that Nvidia can reach up to five times better cost efficiency than AMD at specific output speeds.

SemiAnalysis notes that as production inference traffic shifts toward long-context, multi-turn agentic sessions, the performance gap between accelerators may widen. The study utilized a corpus of over 8,000 sessions and 610 billion tokens to model real-world traffic patterns, including median input lengths of 142,000 tokens.