What Is a Frontier Model?
A frontier model is an AI model at or near the leading edge of capability. The term is used in research and policy, but the threshold varies: some definitions emphasise broad performance, while others focus on capabilities that could create serious risks.
“Frontier” is a relative description, not a certification that a model is AGI, superintelligent or safe.
A moving boundary, not a model architecture
A model can be called frontier because of its performance compared with other systems, not because it uses a special architecture. As methods improve, the boundary can move.
Training cost or scale is sometimes used as a proxy, but it is not equivalent to capability. Better algorithms, data or post-training can change what a system does without making its size alone a reliable guide.
General-purpose capability and specific hazards
A general-purpose model can support many tasks, such as text analysis, coding and work with images. It is not necessarily the strongest system at every task. A specialist model may exceed it in a particular domain.
Policy discussions sometimes focus on frontier systems because their capabilities could enable consequential misuse. This is a question about what the model makes possible, not simply how popular or expensive it is.
The deployed system matters too
A model operating alone and the same model connected to tools can present different risks. Access to a browser, code execution or external services changes the range of possible actions.
Evaluators therefore need to specify the whole setup: instructions, permissions, tools, time and human assistance. Testing a base model without those conditions may not represent a deployed application.
Safety assessments can examine harmful capabilities, robustness and the effectiveness of controls. Results should be read with their limits rather than as a universal verdict.
Questions worth asking
- What definition of frontier is being used?
- Which capabilities were measured, and against what comparison?
- Does the evaluation cover the deployed system or only the underlying model?
- What failures and uncertainties were reported?
These questions help separate a substantive capability claim from a marketing label. The answer may differ across research, regulation and product descriptions.
Sources and further reading
These sources offer definitions, frameworks or arguments relevant to this explanation. They do not imply endorsement of this site.
- Frontier AI: capabilities and risks (UK government, 2023) — one policy use of the term; definitions and capability thresholds vary.
- Levels of AGI (Morris and colleagues, 2023) — a research framework that separates breadth of capability from performance; not a universally agreed definition.