Frontier AI model

At any moment, a handful of general-purpose AI models are the most capable there are. Together they form the frontier. The frontier moves as new models appear. A group of leading AI labs judges models by what they can do. For this group, a model counts as frontier when it beats every model that has been on the market for at least a year. Some US state laws instead include every model trained with more than a fixed amount of computing power. EU law has a similar category, the general-purpose AI model with systemic risk. It sets a lower bar and puts legal duties on the models' makers. None of these definitions looks at how a model is released. In practice it still matters, because the best models with published weights trail the frontier by a few months.