What is a frontier model?

A frontier model is one of the handful of most advanced AI models being trained at any given time — the ones pushing right up against the current edge of what's possible, usually built by the few labs with enough money and computing power to attempt it, like OpenAI, Anthropic, and Google DeepMind.

How is a frontier model different from a foundation model?

That's different from a foundation model, which describes how a model is built: a big base model trained once and then adapted for many different tasks. Frontier model describes where a model sits on the capability ladder right now, not how it was built. Today's top releases from OpenAI, Anthropic, and Google count as frontier models; last year's versions don't, because the frontier keeps moving forward.

In practice a single model can be both at once — a foundation model that also happens to be at the frontier — which is part of why the two terms get mixed up so often.

Why does the term matter for regulation?

The term matters because regulators use it as a trigger for oversight, tied to how much compute went into training:

The EU AI Act applies extra safety and transparency rules above its threshold. California's SB 53 and New York's RAISE Act set the bar even higher and require large developers to publish safety protocols and report serious incidents quickly.

Why do frontier models get the most scrutiny?

The logic behind all of this: the newest, most powerful models are the ones whose risks are least understood at the moment they ship, so they get the most scrutiny. When you see "frontier model" in a headline, read it as shorthand for the most capable AI systems that currently exist, and the ones regulators are watching closest.