What it is. A frontier model is one of the most capable general-purpose AI models of its time, the ones at the leading edge of what AI can do.
An analogy. Think of aircraft. A small propeller plane and a jumbo jet both fly, but only the jumbo carries hundreds of people, so it faces stricter checks before every flight. Frontier models are the jumbo jets of AI. Most AI tools are the small planes.
How do you tell which is which? There is no single definition, so lawmakers use stand-ins. The most common is the amount of computing used to train the model, because more training compute has tended to bring more capability. California’s frontier AI law (SB 53) applies to models trained with more than 1026 computing operations, a 1 followed by 26 zeros. The EU’s AI Act presumes that a general-purpose model trained with more than 1025 operations poses “systemic risk” and owes extra duties. Compute is a rough guide, not a perfect one: a smaller, cleverly trained model can sometimes match a bigger one.
Why it showed up in the news this week. Two announcements aimed squarely at this group. On 9 November six governments and ten AI developers signed a joint declaration in Paris committing to shared ways of testing frontier models for misuse in cyber attacks and weapons, and for loss of control. On 10 November Australia said companies building the most advanced models must meet six mandatory obligations, including independent evaluation and rapid incident reporting, enforced by a new regulator.
Why only the frontier? These models are general: the same system that writes a poem can help write code, including harmful code. Regulators argue the biggest risks sit with the most capable systems, so focusing there avoids burying a small app maker in paperwork.
One thing to remember. “Frontier” is a moving line. Today’s frontier model is next year’s ordinary one, which is why some laws let officials update their thresholds over time.
Sources: Joint Declaration on the Paris Call for AI Safety and Security, 9 Nov 2026, newswire.ca; Prime Minister of Australia, “New Australian AI safety and security protections”, 10 Nov 2026, pm.gov.au; Future of Privacy Forum on SB 53 thresholds, fpf.org; Regulation (EU) 2024/1689 (AI Act), Article 51, eur-lex.europa.eu.