Frontier model

In plain English

The frontier moves. A model that was frontier two years ago is often matched by much cheaper systems today. Labs use the term for their newest flagship systems, and governments use it to aim safety rules at the few developers able to train them, usually by a training compute threshold or by revenue. California's SB 53 and New York's RAISE Act, for example, cover developers that train models above 10^26 operations, with the heaviest duties for those with more than $500 million in annual revenue.

Frontier models are where new and dangerous capabilities appear first, from advanced cyber offense to long running autonomous agents. They are also where most safety frameworks, outside testing and incident reporting focus. Critics note that thresholds based on compute can miss capable models trained more efficiently, and that open weight releases bring near frontier capability to anyone who downloads them.

Why it matters for rogue SI

Rogue SI, if it comes, will come from the frontier. The incidents with the highest Rogue Scale ratings in 2026 involve frontier models tested with real tools and real network access.

Laws and rules that use it

• Voluntary or guidance AI Action Plan United States
• Proposed FRONTIER Act United States
• In force SB 53 California
• In force EO N-9-26 California
• Passed, not yet in force RAISE Act New York
• Passed, not yet in force SB 315 Illinois
• Passed, not yet in force SB 26-189 Colorado
• Voluntary or guidance UK AISI United Kingdom

Sources

1. White House Unveils America's AI Action Plan The White House · Jul 23, 2025
2. H.R. 9925, FRONTIER Act, 119th Congress LegiScan