In plain terms
The frontier is the line between what AI can and cannot yet do, and frontier models are the handful that sit on it: the newest flagship models of the leading labs. The label is relative and temporary. The frontier model of two years ago is matched today by models that are far cheaper, smaller or free to download.
Why it matters
Two practical consequences follow. For buyers: frontier capability costs the most and is needed less often than suppliers suggest. Many tasks are handled just as well by a model one tier down, and a task that needs a frontier model today will be cheap to run in a year or two. Build so that models can be swapped. For policy: because their abilities are new and not fully understood, frontier models are where safety testing and regulation concentrate.
Example
A consultancy builds a contract-review tool on the most capable model available, at 18 dollars per contract. A year later a mid-tier model from the same provider passes the same test set with equal accuracy at 2 dollars. Because the team had kept its test set and had not tied the tool to one model, the switch takes an afternoon.
Most often confused with
Frontier Model vs. Foundation model
All frontier models are foundation models; few foundation models are frontier models. “Foundation” describes a role, and stays true as a model ages. “Frontier” describes a rank, and expires.
Origin: The term spread through AI policy discussions in 2023, the year in which several leading labs founded the Frontier Model Forum.
Under the hood
There is no fixed threshold. Working indicators: leading results on demanding benchmarks (long software tasks, scientific reasoning, agentic work), the computing power used in training, and release as a lab's flagship. Developers of these models publish safety frameworks that tie stronger safeguards to capability thresholds, and run evaluations before release for risks such as help with biological or cyber attacks and autonomous behaviour. In regulation, the EU AI Act presumes systemic risk for general-purpose models trained with more than 10^25 floating-point operations, which brings additional obligations; other jurisdictions use comparable tests based on compute or capability. Economics: frontier capability is expensive at release, and equivalent capability has repeatedly become far cheaper within one to two years as smaller and open-weight models catch up.