SemiAnalysis put real coding-agent traffic through an open benchmark. On one GLM 5.3 configuration Nvidia runs 5x cheaper than AMD, and the gap sits in serving software.

SemiAnalysis put real coding-agent traffic through an open benchmark. On one GLM 5.3 configuration Nvidia runs 5x cheaper than AMD, and the gap sits in serving software.

A new benchmark built from real coding-agent sessions gives Nvidia a major cost-efficiency advantage over AMD—and suggests software, not just chips, is widening the gap.