What is a frontier model?

"The fastest car in the world" names a different car every few years. Nobody thinks it's a kind of engine. It's a title, and someone always comes along to take it. "Frontier model" works the same way. You'll see it in news stories and in government papers about artificial intelligence (AI), and it means one of the most capable AI models at a given moment.

The word describes where a model ranks, rather than how it's built. There's no special part inside that makes a model frontier. It just sits at the leading edge of what models can do right now. When the United Kingdom government prepared for an international AI safety meeting in 2023, it described frontier AI as highly capable general-purpose models that match or exceed the most advanced models of the day. Notice those last words. The definition points at today, so it moves with today.³

Building one takes huge amounts of computing power, data, and money. So frontier models usually come from a small number of large companies, which is why you keep hearing the same few names. Stanford University's AI Index for 2026 reports that more than nine in ten notable models released in 2025 came from industry, with universities producing far fewer.¹

Frontier is a moving target. A model that leads the field this spring can look ordinary by next year, once newer ones pass it. The older model doesn't get worse. It's just as good at what it did. The edge moved on without it. So when a story calls something frontier, check the date. A two-year-old article about frontier models is probably describing models that have since been passed.

Governments often aim their newest AI rules at the most capable systems first, because that's where they expect new risks to show up. The European Union's AI Act, in force since 2024, is one example. It uses the amount of computing that went into training as a first test, and asks the makers of the most powerful general-purpose models for extra safety testing and for reports when something serious goes wrong. Other rules, in other places, can draw the line differently. So the word in a law may not mean exactly what you'd guess from a headline.⁴

The race at the top has also tightened. Stanford's AI Index for 2026 reports that the leading models from several companies now sit close together on a popular ranking where people vote between two answers. It also found the best freely downloadable models only slightly behind the best closed ones, a gap that had nearly closed in 2024 and widened again in 2025. When the leaders are that close, "the most capable model" can depend on which test you pick.¹²

Most everyday jobs, like drafting an email, summarizing a page, or brainstorming names for a project, don't need the frontier. Models a step or two behind it often handle them well, for less money and in less time. So the next time a product calls itself frontier, try a plainer question. Frontier at what, and as of when?