Frontier Model
Most capable AI models requiring enhanced oversight.
How the laws define this
Regulators consistently frame frontier models as highly capable, advanced AI systems whose power and scale create significant potential risks to public safety, national security, or critical infrastructure. Despite agreeing on this potential for severe societal impact, these authorities diverge on whether to classify a model using strict quantitative benchmarks or broader capability assessments. California SB 53 — Transparency in Frontier Artificial Intelligence Act (TFAIA) relies on a precise technical cutoff of more than 10^26 computing operations, while the Responsible AI Safety & Education Act focuses on high computational power and financial investment during training. Furthermore, the Transparency in Frontier Artificial Intelligence Act highlights that the California Department of Technology annually reviews quantitative thresholds. By contrast, instruments such as the Artificial Intelligence Safety Measures Act and Safe and Responsible AI in Australia — Government Interim Response to Consultation define frontier models through qualitative risk profiles, emphasizing general-purpose deployment, emergent properties, or catastrophic threats.
Synthesised from the 8 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| California, United States | California SB 53 — Transparency in Frontier Artificial Intelligence Act (TFAIA) | In Force | Defines frontier model by a precise quantitative threshold of over 10^26 computing operations, including fine-tuning or modifications. |
| New York, United States | An act to amend the general business law, in relation to transparency and safety requirements for developers of artificial intelligence frontier models; to amend a chapter of the laws of 2025 amending the general business law relating to the training and use of artificial intelligence frontier models, as proposed in legislative bills numbers S. 6953-B and A. 6453-B, in relation to the effectiveness thereof; and to repeal certain provisions of the general business law, relating thereto | Awaiting Entry | Emphasizes sophisticated AI systems posing significant risks to public safety or well-being due to their power or complexity. |
| Illinois, United States | Artificial Intelligence Safety Measures Act | Adopted | Limits to highly capable models posing severe or catastrophic risks to public safety, national security, or critical infrastructure. |
| United States | The Great American AI Act | Proposed | Highlights most advanced and powerful AI systems with significant capabilities, systemic impact, and substantial computational resources. |
| Australia | Safe and Responsible AI in Australia — Government Interim Response to Consultation | Draft | Focuses on very large or general-purpose models deployable across sectors, scalable rapidly, and posing unique systemic risks. |
| California, United States | Transparency in Frontier Artificial Intelligence Act | In Force | Describes most advanced general-purpose systems with annually reviewed quantitative thresholds by the California Department of Technology. |
| California, United States | California Executive Order N-9-26 — Independent Oversight of Frontier AI | In Force | Centers on leading-edge systems posing heightened safety, security, or misuse concerns due to their capabilities. |
| New York, United States | Responsible AI Safety & Education Act | Awaiting Entry | Includes advanced models with significant computational power and financial investment, or smaller derived models with similar capabilities. |
| United States | TRUMP AMERICA AI Act | Draft | Specifies advanced systems demonstrating emergent properties or high computational power with potential national security or economic impacts. |
Definitions (9)
A foundation model trained using more than 10^26 computing operations, counting the original training run plus any fine‑tuning or modifications; this threshold identifies the largest, most resource‑intensive AI systems subject to the law.
Refers to sophisticated artificial intelligence systems that possess or could acquire capabilities that, due to their power or complexity, pose significant risks to public safety or well-being. These models are the primary focus of the Act's regulatory framework.
A highly capable artificial intelligence model that, due to its advanced capabilities, could potentially pose severe or catastrophic risks to public safety, national security, or critical infrastructure. These models are typically at the leading edge of AI development.
Refers to the most advanced and powerful AI systems, characterized by their significant capabilities and potential for systemic impact, often requiring substantial computational resources for training and deployment.
A frontier model refers to the most advanced and general-purpose AI systems, such as those developed by leading AI firms. The California Department of Technology annually reviews the specific quantitative thresholds for these models to ensure the definition remains current with technological advancements.
Very large or general-purpose AI models that can be embedded across many settings and scale rapidly; they may present unique systemic risks due to their generality and broad cross-sector deployment. The document signals tailored safety mechanisms and possible model-specific obligations for such systems.
These are advanced AI models characterized by significant computational power and financial investment during their training phase, or smaller models derived from them with similar capabilities. They are identified as having the capacity to cause significant societal impact.
Advanced AI systems that demonstrate emergent properties or high computational power, potentially impacting national security or economic stability.
A frontier model is a leading-edge AI system that can pose heightened safety, security, or misuse concerns because of its capabilities. The order treats frontier models as the core subject of verification, incident reporting, and possible future control measures.
Not to be confused with
Frontier AI model refers specifically to general-purpose models near cutting-edge performance, while frontier model encompasses broader definitions of advanced AI systems across regulations.
Foundation models / Frontier AI denotes highly capable general-purpose models with broad applicability, whereas frontier model often includes specific risk or compute thresholds not always present there.
Questions and answers
What compute threshold defines a frontier model in California law?
California SB 53 defines it as a foundation model trained using more than 10^26 computing operations, including original training, fine-tuning, or modifications.
Which US states require audits or safety frameworks for frontier models?
California's SB 53 and TFAIA require safety frameworks and incident reporting, while Illinois' Artificial Intelligence Safety Measures Act mandates independent third-party audits.
How do frontier model definitions vary across proposed US bills?
Bills like the Great American AI Act and TRUMP AMERICA AI Act describe them as advanced systems with significant capabilities, systemic impacts, emergent properties, or high compute power.
Related Terms
Frontier AI model
General-purpose models near cutting‑edge performance....
Foundation models / Frontier AI
Highly capable, general-purpose AI models with broad applicability....
Frontier models
Very large or general-purpose models with cross-sector impact....
Model
A trained algorithmic artifact that performs tasks using learned parameters....
Foundation Models
Large-scale models trained on broad data for many tasks....