Frontier model
A term for the most-capable AI models available at any given time — those pushing the leading edge of capability, typically trained with the most compute.
"Frontier model" is a semi-formal term for AI models at the leading edge of capability. In 2026 the frontier includes Claude Sonnet 4.6, Opus 4.6, GPT-5.6 Sol, Gemini 3.1 Pro, and a few others. The set changes every few months as new releases push the boundary.
The label matters for two reasons. First, pricing: frontier models cost more to train (typically hundreds of millions of dollars) and more to run per query. That cost flows through to what you pay for consumer products or API access. Second, regulation: the EU AI Act and several US state laws use compute-based thresholds (typically 10²⁵ FLOP of training compute) to define which models are subject to additional requirements. Frontier models sit above those thresholds.
The label doesn't have a strict definition. In casual use, "frontier" and "state-of-the-art" and "flagship" get used interchangeably. In policy contexts, the compute threshold is the operative definition.