A frontier model is an AI system, almost always a large foundation model, that represents the current cutting edge of capability, typically exceeding the performance of any previously widely deployed model on a broad range of tasks. Unlike more technical terms such as LLM or Transformer (architecture), frontier model is primarily a policy and governance term, used to identify which systems warrant heightened scrutiny because their capabilities are advancing faster than institutions can fully evaluate or regulate them.
Origins of the term
The term gained currency around 2023 as labs including OpenAI, Anthropic, and Google DeepMind formed the Frontier Model Forum, an industry body focused on safety research and best practices for the most capable models. Policy documents, including work from the UK's AI Safety Institute and the Bletchley Declaration following the 2023 AI Safety Summit, adopted similar language to scope which systems international coordination efforts should prioritize, generally defined by very large training compute or demonstrated capabilities that could pose risks in domains like cybersecurity or biological weapons design.
Characteristics
Frontier models are typically defined less by a fixed technical threshold than by relative position: a model counts as frontier if it is at or near the top of current capability across general tasks, as measured informally by benchmarks and community evaluation platforms such as LMArena. Frontier status is inherently temporary, since a model that was frontier on release is typically surpassed within months by a competitor or successor; the label describes a model's position at a point in time rather than a permanent category. Frontier labs, the organizations that train these systems, include OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, and increasingly Chinese labs such as DeepSeek and Alibaba's Qwen team.
Policy relevance
Because frontier models are the first to exhibit new capabilities, they are the focus of voluntary and regulatory commitments around pre-deployment testing, such as the responsible scaling policies adopted by several labs, which tie additional safety measures to capability thresholds. The EU AI Act similarly created special obligations for general-purpose AI models with systemic risk, operationalized in part through a compute threshold, reflecting the same underlying concern as the frontier model concept: that a small number of unusually capable systems warrant scrutiny beyond what applies to smaller or narrower AI products.
Debate
Critics of the term argue it can function as a marketing label, implying leadership or safety-consciousness that a lab's actual practices may not match, and that defining risk by capability alone overlooks how a model is deployed and by whom. Others note that focusing regulatory attention narrowly on frontier models may leave the far larger number of applications built on slightly-behind or open-weights models under-scrutinized, even though those systems can pose similar risks once widely deployed. Despite the debate, the term remains standard shorthand in AI governance discussions as of 2025.