Nvidia is widening its performance and cost-efficiency lead over AMD in a new benchmark designed specifically for the way modern AI agents operate. The AgentX benchmark, released by SemiAnalysis, uses real-world coding-agent sessions rather than traditional fixed-prompt tests, offering a different perspective on the hardware competition between Nvidia and AMD.
A Benchmark Built for AI Agents
AgentX focuses on long-context, multi-turn workloads that more closely resemble production AI-agent traffic. Instead of simply measuring how quickly a system responds to isolated prompts, the benchmark replays actual coding-agent sessions involving repeated interactions, large contexts and periods of inactivity between turns.
SemiAnalysis created its dataset from more than 8,000 internal sessions containing roughly 610 billion tokens. The public version 1.0 includes 393 anonymized Claude Code sessions, with the data transformed to preserve important caching relationships without exposing the original prompts or code.
That distinction matters because AI agents behave differently from conventional chatbots. They can maintain long conversations, repeatedly access the same context and generate relatively small amounts of output after processing very large inputs.
Nvidia Shows a Major Cost Advantage
The results point to a substantial advantage for Nvidia in certain agentic inference configurations.
According to SemiAnalysis, Nvidia hardware running GLM 5.3 through the open-source SGLang stack achieved up to five times better cost efficiency than AMD at an operating point of 150 output tokens per second per user.
However, the figure should not be interpreted as Nvidia being five times faster than AMD across all AI workloads. The comparison depends heavily on the model, software stack, throughput target and infrastructure configuration. SemiAnalysis’ own comparisons show the advantage changing significantly as interactivity requirements increase.
Software Is Becoming a Competitive Weapon
One of the most important findings from AgentX is that the hardware itself may not tell the whole story.
Nvidia’s advantage is partly attributed to its mature software ecosystem, including sophisticated cache management, request routing and incremental tokenization. These technologies become particularly valuable when AI systems repeatedly process large amounts of context across multiple turns.
AMD has been developing its own software infrastructure, including the ATOM inference engine. But the relatively limited adoption of that stack means Nvidia can potentially extract greater efficiency from its hardware through years of software optimization.
