An AI model at or near the cutting edge of capabilities, typically characterized by training compute, parameter count, and emergent abilities that may pose novel risks not present in less capable systems. Regulatory frameworks have attempted to formalize this concept using compute thresholds: President Biden's EO 14110 defined "dual-use foundation model" at 10²⁶ FLOPs; the EU AI Act presumes "systemic risk" for GPAI models trained above 10²⁵ FLOPs. Though EO 14110 was rescinded, compute thresholds remain relevant to export controls, international coordination, and voluntary industry commitments. The threshold for "frontier" shifts continuously—today's frontier becomes tomorrow's baseline.
See: Dual-use foundation model; Executive Orders on AI; Export controls; Foundation model; GPAI; Systemic risk