Models
Frontier Model
Also known as: frontier AI
The most capable, cutting-edge AI models at any given time, typically the largest and most expensive, built by a handful of leading labs.
A frontier model is a top-of-the-line AI, the newest, most capable model at any moment, usually one of a few built by the leading AI labs. These are the models that push what AI can do and set the headlines. Because they are so powerful, expensive to build, and hard to predict, they get extra attention from both researchers and regulators concerned with safety.
"Frontier model" refers to the most advanced AI models of the day, typically very large foundation models at or beyond the current state of the art. The label is relative and moves over time as new models arrive.
Why the term exists
- Capability: they define what is currently possible.
- Cost and concentration: training them takes enormous compute and money, so only a few well-resourced labs build them.
- Safety and policy: because their capabilities (and risks) are hard to predict in advance, "frontier AI" has become a focus of governance and voluntary safety commitments.
"Frontier model" designates the current capability frontier, typically the largest, most compute-intensive foundation models exhibiting the strongest and least-predictable capabilities, a deliberately relative and policy-salient term (Anderljung et al. 2023). Its significance is threefold: it marks the research state of the art driven by scaling and method advances; it reflects extreme concentration, only a handful of labs can afford frontier-scale compute; and it anchors governance discourse.
Because frontier systems can display emergent and dual-use capabilities not evident pre-training, they are the focus of dangerous-capability evaluations, responsible-scaling/pre-deployment testing, and regulatory attention (e.g. GPAI-with-systemic-risk provisions in the EU AI Act, safety-institute evaluations). The concept is contested at the edges, capability is multidimensional and "frontier" is not sharply defined, but it usefully separates the highest-risk, highest-capability tier from the broad ecosystem of smaller and open models.