frontier model

In depth

A relative definition

The definition is relative, and that is what makes it unusual among the categories of this floor. It designates the models that push back the limit of what is achievable, the ones whose capabilities had not yet been reached. No value fixes it: not a number of parameters, not a compute budget, not a score. Two properties of shape follow. The frontier is a zone occupied simultaneously by a small number of models, coming from a small number of organisations, and its membership keeps turning over. A model enters it because it does what no other did, and leaves it once that move has become commonplace: the category therefore qualifies a model by its neighbourhood, not by its contents.

A trigger

What gives the term its weight is that it serves as a trigger. The safety policies that organisations set themselves, like the texts that regulate the field, use it to attach obligations to a very small number of models: evaluations of dangerous capabilities before release, documentation, incident reporting, monitoring after deployment. The reasoning is the same everywhere: rather than regulating every model in the same way, you regulate those whose capabilities could produce a risk at the scale of society, accepting that the list will move. Since a capability is hard to measure and easy to dispute, those texts rarely rest on it and take an indirect proxy instead, the most common being the volume of compute spent during training, together with a revisable threshold. That is not a definition, it is an operational approximation: it has the merit of being verifiable, and the flaw of ageing fast, since progress in training methods lowers the compute needed for a given capability. For an organisation that buys, the word mainly signals a regime: more documentation, more announced checks, and longer lead times before release.

Three traps

Three traps are worth naming. The first is commercial: nobody protects the term, every organisation that publishes a model has an interest in calling it frontier, and the promotional use of the word has long outgrown its descriptive use. The second is to read it as a level of quality for your own use: the frontier shows on the hardest tasks, the ones where other models fail, and most professional uses are not among them. A frontier model costs more there and answers more slowly for an equivalent result, and it is an evaluation set of your own, never a public ranking, that settles the question. The third is the confusion with the benchmark, whose tests saturate and become contaminated, so that a place on one says little about the real position. One structural consequence remains: frontier models are almost always served remotely and their weights stay closed, because published weights cannot be withdrawn and cannot be monitored, which makes the obligations attached to the status unenforceable. The contrast with open weight models is therefore no accident, and the gap between the two is measured in months rather than in generations.