The frontier is the leading edge of what AI can currently do. A frontier model is a model at that edge: the newest, most capable tier, made by one of the few companies with the money and computing power to train them, such as Anthropic, OpenAI and Google.
The important thing about the frontier is that it moves. Today's frontier model becomes next year's ordinary one, and the price for that level of capability keeps falling. What required the most expensive model eighteen months ago often runs fine on a cheap fast one today.
Why you keep hearing it
Because it is shorthand in every serious AI discussion. "Frontier capabilities" means what the best models can newly do. "Frontier labs" means the companies making them. Products also borrow the word for marketing, since "powered by a frontier model" sounds better than "we use the same model as everyone else."
What it means for you
Two practical points. First, when you read about what AI can or cannot do, check which model the claim is about. A lot of confident "AI can't do this" statements are based on older or smaller models. Second, you rarely need the frontier for everyday marketing work. The expensive newest model is for the hardest problems; most tasks run well on the tier below, at a fraction of the cost.