Frontier Model

“Frontier model” is a relative label, not a fixed tier. It points at the handful of most capable general models at the current moment — the latest flagship releases that define the state of the art. Because the field moves fast, the frontier keeps shifting: a model that was frontier a year ago is now a mid-tier option you can often replace with something smaller and cheaper. The Frontier Model Forum, an industry body founded by several of the labs, describes its scope the same way: the general-purpose models that make up the state of the art, a set that shifts as the field progresses.

In policy writing the term usually carries a second meaning. Regulation papers such as Anderljung and colleagues’ Frontier AI Regulation use “frontier AI” for highly capable models that could have capabilities dangerous enough to threaten public safety, so a “frontier” label in a law or a lab’s safety framework is about risk thresholds as well as capability.

The term matters for build decisions. Frontier models earn their cost on open-ended, hard reasoning where capability is the constraint. For narrow, repeated tasks, you usually don’t need the frontier — the smart move is the smallest model that still passes your evals, reserving the frontier for the work that requires it.