The phrase appears throughout AI policy, in export control discussions, in national security testimony, and across every accusation made during 2026. It has no agreed technical definition.
The working meaning
A frontier model is one at the leading edge of current capability: among the largest, most capable and most general systems that exist at a given moment.
Note what that depends on. It is relative, defined by what else exists. It is temporal, so today's frontier model is next year's mid-tier option. And it is capability-based, which makes it awkward to legislate against, because capability is difficult to measure and easy to dispute.
Frontier, foundation, general-purpose
- Foundation model: trained broadly on large data, adaptable to many downstream tasks. A size and generality claim, not a quality claim.
- General-purpose AI model: the term the EU AI Act uses, similar in scope, with legal consequences attached.
- Frontier model: a foundation model currently among the most capable in existence.
Every frontier model is a foundation model. The overwhelming majority of foundation models are not frontier models.
Why regulators reach for compute
Since capability is contested, policy has largely settled on training compute as a proxy. It has one decisive advantage: it is measurable, and to some degree verifiable through hardware supply chains.
The EU AI Act uses a threshold of 10 to the power of 25 floating point operations of training compute as a presumption of systemic risk for general-purpose AI models, which triggers additional obligations. Other jurisdictions have used or proposed thresholds at similar orders of magnitude, and US policy in this area has shifted more than once.
Why the compute proxy is weakening
The proxy assumes capability tracks training compute. That relationship is loosening, and distillation is one of the reasons.
A distilled model can approach its teacher's capability while using a small fraction of the compute the teacher required. Under a compute-threshold regime it may fall well below the line while behaving like something well above it. The threshold measures how a model was made, not what it can do.
This is the structural problem running underneath the whole 2026 dispute. Export controls and compute thresholds regulate the manufacture of capability. Distillation transfers capability without repeating the manufacture. We examine the consequences in The cascade problem.
Why the label matters commercially
Being called a frontier model attracts obligations: evaluation requirements, reporting, and in some proposals licensing. It also attracts attention, which is why the term appears so readily in marketing and so reluctantly in regulatory filings.
When you encounter the phrase, it is usually worth asking which sense is intended: a claim about capability, a claim about compute, or a claim about which rules apply.
Common questions
What is a frontier AI model?
A frontier model is one at or near the leading edge of current capability, typically among the largest and most capable general-purpose systems available at a given moment. It is a relative term rather than a fixed technical category, so what qualifies changes continuously.
What is the difference between a foundation model and a frontier model?
A foundation model is any large model trained broadly and adapted to many downstream tasks. A frontier model is a foundation model that is currently among the most capable in existence. All frontier models are foundation models; most foundation models are not frontier.
How do regulators define frontier models?
Mostly by training compute, because it is measurable and verifiable in a way capability is not. The EU AI Act uses a threshold of 10 to the 25 floating point operations to presume systemic risk for general-purpose AI models. Compute is a proxy rather than a definition, and its usefulness declines as efficiency improves.