Frontier AI is a policy and technical term used for the most capable general-purpose artificial-intelligence systems near the cutting edge of current performance.
The definition is intentionally relative. UK government documents used for international AI-safety work describe frontier AI as highly capable general-purpose models or systems that can perform a wide variety of tasks and match or exceed the capabilities of today's most advanced models. As technology advances, yesterday's frontier may become ordinary.
Governments focus on frontier systems because higher capability can create both larger benefits and larger risks. A very capable model may help with science, software, education and productivity, but the same broad capability can raise concerns about cyber misuse, biological or chemical assistance, deception, autonomous action or loss of control.
That does not mean every frontier model is dangerous. Risk depends on what a system can actually do, how it is deployed, what safeguards exist and who can access it.
This is why many safety frameworks focus on evaluations and thresholds rather than model size alone. Developers may test capabilities before deployment, assess misuse risks, strengthen access controls and publish safety information.
Another important point is that there is no single global legal definition. Different governments and institutions may use slightly different language, and the boundary changes over time.
So when a company or government calls a system frontier AI, readers should ask: which capabilities put it at the frontier, what evidence supports those claims, what risks were tested, and what safeguards are in place?