Frontier Model

A Frontier Model is a Machine Learning model at or beyond the current state of the art in capabilities, generally a high-capacity General-Purpose AI system able to perform complex tasks across diverse domains. The designation is relative and time-dependent; historical examples include GPT-4 by OpenAI, the Claude 3 family by Anthropic, and early Gemini models by Google DeepMind. These systems are often Large Language Models with many Parameters and substantial Compute requirements, commonly involving GPU clusters from suppliers such as NVIDIA, but frontier status is not defined by model type, parameter count, a fixed compute threshold, hardware vendor, or the presence of Emergent Abilities. Their potential systemic impacts and associated risks make them a major focus of many AI Safety initiatives—including work on AI Alignment, Red Teaming, and responsible deployment—and relevant to some regulatory provisions. Regulation also addresses other general-purpose and High-Risk AI systems: the EU AI Act, for example, imposes additional obligations on general-purpose models presenting Systemic Risk, rather than using “frontier model” as its general legal category. The Frontier Model Forum is an industry-supported nonprofit that advances safety research, Model Evaluation, risk identification and mitigation, information sharing, and Red Teaming guidance; its work can inform practices and standards, but it is not a regulator or a body that establishes binding international standards.