Frontier model

Frontier model is the term for the most advanced general-purpose systems at any given moment — the handful of models that push the boundary of what AI can do. They are trained on enormous datasets using vast amounts of computing power, at a cost that runs into the hundreds of millions, which is why only a small number of well-funded labs build them.

The label matters beyond bragging rights. Because these models are the most capable, they are also where the most serious risks and the most valuable capabilities first appear. That has made them the focus of AI safety research and of government attention. Britain’s AI Security Institute (formerly the AI Safety Institute) was set up largely to test frontier models before and after release — work that depends entirely on getting access to them, as the UK found when that access to US models was switched off for a fortnight in 2026. The UK’s own frontier capability is anchored by Google DeepMind in London, and not depending wholly on American models is a large part of the country’s sovereign-AI push.

The frontier is a moving target. Today’s frontier model is next year’s ordinary one, as techniques spread and computing gets cheaper. The word describes a position at the leading edge, not a fixed level of ability.