What Is a Frontier Model?

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.

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.

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.

• 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.

AGI concerns broad, flexible intelligence, often at human level. Superintelligence would exceed the best humans across virtually all relevant cognitive domains. “Frontier” merely locates a system near the leading edge under some definition.

A model can be frontier without meeting either of those thresholds. Equally, calling a system frontier says nothing by itself about consciousness, judgement or reliability.