Frontier Model
A model at or beyond the current state of the art in general capability — the ones regulators watch most closely.
Frontier model has no crisp technical threshold; it means whatever is currently most capable and therefore least well understood. The term matters mainly in policy, where extra evaluation and reporting obligations attach to the leading edge. The EU AI Act’s nearest concept is a general-purpose AI model with systemic risk.
In practice: Every generation, yesterday’s frontier model becomes today’s ordinary default.