Frontier models

A frontier model is simply the best a lab can build at a given moment: trained at the limit of available compute, data, and engineering. The label is relative and shifts with each major release. It matters because these models set the capability ceiling that everything else — benchmarks, safety research, commercial products — is measured against. The hard part is that building them requires billions of dollars, months of training time, and proprietary infrastructure that only a handful of organizations can assemble, which concentrates who decides what the frontier looks like.