What it is. "Frontier model" is the working term, used by major AI labs and a growing number of regulators, for the largest and most capable general-purpose AI models at any given time. GPT-5, Claude Opus 4, Gemini 3, and Llama 4 sit at the frontier today. The smaller, distilled, or open-weights versions of those models (a 7B parameter Llama, for example) do not.
Why the term exists. Regulators and labs needed a label that captures "the small set of models that pose novel risks because of their capability level" without freezing a number into law. A pure parameter-count threshold (say, 70 billion) ages badly because capability rises with training data and post-training even when parameter count stays flat. The term "frontier model" lets policy attach to capability instead of architecture.
The working definition, roughly. Most regulatory drafts (the U.S. AI Executive Order, the EU AI Act's general-purpose AI obligations, the UK AI Safety Institute's framing) anchor on training compute as the proxy. The 1025 to 1026 FLOPs range tends to be where "frontier" starts. A few add a capability-test fallback: a model can be designated frontier if it crosses certain benchmarks regardless of training compute. The exact thresholds are in active flux and tend to drift up as compute gets cheaper.
Why operators should care. If you're a model provider, "frontier" status triggers reporting requirements: training runs above the threshold may need to be disclosed to the relevant government, and red-team evaluations may need to be filed. If you're an enterprise buyer, frontier-classified models are the ones most likely to face export restrictions (especially to China), capability-based deployment carve-outs, and the most aggressive safety testing requirements before each release. Pricing and roadmap commitments at the frontier are also less stable than at the tier below.
What it does not mean. "Frontier" is not a quality endorsement. A frontier model isn't necessarily the best fit for any specific task. For most production workloads, smaller and cheaper models, often distilled from frontier models, win on cost and latency.
Bottom Line
When a politician, a lab safety post, or an export control rule says "frontier model," they mean the handful of largest models from a handful of labs, not the open-source one you're running on a single GPU.