Frontier Model

"Frontier" describes position, not quality ranking. A frontier model is one operating at the edge of what has been demonstrated — trained at a scale only a handful of organizations can afford, and capable of things no previously released model could do. The labs producing them are called frontier labs for the same reason.

The practical definition has become partly regulatory. Several frameworks now define the category by training compute above a threshold, on the theory that capability tracks scale closely enough for compute to serve as a workable proxy. That gives the term a legal meaning it did not originally have, and it is why you now see compute thresholds quoted in policy documents.

The set is small and changes hands. Frontier status is a claim about a moment, not a permanent property of a company.

Most AI regulation that targets models rather than uses targets this category specifically — obligations attach above a compute threshold, so whether a model is "frontier" determines which rules apply to whoever built it. For buyers, it is also a rough guide to cost: frontier models are the expensive tier, and a great deal of practical engineering is about not using one where a smaller model would do.

That "frontier" means "best". It means at the capability edge, from a lab running the largest training runs. Those usually correlate, but the claim is about scale and recency rather than about being the top scorer on your task. A smaller, cheaper, older model routinely beats a frontier model on a narrow job it was tuned for.

That it is a fixed list of companies. Frontier status describes a moment. Labs enter and leave the set as releases land, and a model that was frontier at launch is ordinary within a year or two. Treating it as a permanent brand attribute is how a vendor keeps charging frontier prices for a model that no longer is one.

That the compute threshold is a measure of danger. It is a proxy chosen because it is measurable and hard to fake, not because compute causes harm. A model just under the line is not safe by virtue of being under it, and one just over is not dangerous by virtue of being over. The threshold is a regulatory instrument, not a finding.

Use the term precisely, or avoid it. When comparing options, ask what a model does on your task rather than which tier it belongs to — and default to the cheapest model that passes your eval rather than the most advanced one available.