Token Efficiency Is Frontier Performance

Inference: Frontier model differentiation increasingly includes useful work per generated token alongside capability. Grok 4.5 and GPT-5.6 are both positioned around agentic work with reduced output-token use, creating a concrete basis for workload-level efficiency metrics that combine completion quality, latency, and spend without declaring benchmarks obsolete.