What makes a model a frontier model?
Scale and recency. Frontier models are the largest, newest systems from the leading labs, typically multimodal, typically reasoning models, and typically where new capabilities appear first, often without having been trained for explicitly. The set has no fixed membership: it is whatever currently tops the capability charts, proprietary or open weight, and it turns over in months.
Frontier model vs foundation model: why do both terms exist?
Because they answer different questions. Foundation model came out of academia and describes a fact about usage: a large pre-trained model that downstream systems adapt, which stays true of a model forever. Frontier model came out of policy and safety debates and describes risk: the models capable enough that new behaviors, good and bad, tend to appear in them first, which is why governance regimes such as the EU AI Act attach extra obligations to the most capable general-purpose models specifically. When a vendor says frontier, they are making a claim with an expiry date; when they say foundation, they are telling you where the model sits in the stack.
Why does the frontier keep moving?
Diffusion is the mechanism. Techniques debut at the frontier, then reappear within months in smaller and open models as recipes spread, researchers move between labs, and distilled versions ship. Capability that needed a record-setting training run becomes a commodity behavior a release cycle or two later. For a builder this has one practical reading: the frontier premium buys earliness, not permanence, so pay it only where being early matters. A model routing layer turns that into a per-task decision instead of a per-system one, sending the genuinely hard step to the current frontier and everything else to whatever the last frontier has become.
Should you build on frontier models or cheaper ones?
Both, deliberately. Frontier capability is worth paying for on the steps that fail without it, usually planning and hard reasoning; routine extraction and drafting run as well on a mid-tier LLM or a small language model at a fraction of the price, sometimes one produced from the larger model by distillation. Design for the moving frontier: keep the model swappable, because a system pinned to one lab’s current best inherits that model’s retirement date.