Why the definition matters

Frontier status decides which models carry the heaviest obligations and the most scrutiny. Most definitions use training compute as a proxy for capability because it can be measured before release, although it is an imperfect proxy and efficiency gains make it less reliable each year.

  • EU AI Act: a general-purpose AI model is presumed to have systemic risk above 10^25 floating-point operations, or can be designated by the Commission.
  • California SB 53 (in force 1 January 2026): a frontier model is trained with more than 10^26 operations, counting later fine-tuning; a large frontier developer also has annual revenue above US$500 million.
  • New York RAISE Act (from 1 January 2027): broadly aligned with SB 53 after its 2026 amendment.

The International AI Safety Report, whose second edition was published in February 2026, uses general-purpose AI and speaks of risks at the frontier rather than defining frontier AI.

What frontier developers must do

Under SB 53, large frontier developers must publish a frontier AI framework explaining how they assess and manage catastrophic risk, review it at least annually, publish a transparency report when they release a new frontier model, and report critical safety incidents to the California Office of Emergency Services within 15 days. New York's RAISE Act adds reporting to the state's financial regulator within 72 hours. Under the EU AI Act, providers of systemic-risk models must evaluate and red-team them and report serious incidents to the AI Office. Most major developers also publish voluntary frontier safety frameworks and submit models to pre-release testing by the UK AI Security Institute and the US Center for AI Standards and Innovation.

What it means for an organisation that uses one

Almost no organisation builds a frontier model, but many use one, through an API or inside a vendor's product. The frontier laws put obligations on the developer, not on you, and they produce documents you can use: the developer's framework, its transparency report and the system card for each model. Read them for what was tested and what was not, and record the model version you depend on, because frontier models change often. The capability that makes a model frontier is also what makes it useful for misuse, so the ordinary controls matter more, not less; see third-party risk and AI incident.

Frequently asked questions

What is frontier AI? The most capable general-purpose AI models at or near the leading edge, usually defined by the compute used to train them.

What compute threshold defines a frontier model? It depends on the law: 10^25 operations for the EU AI Act's systemic-risk presumption, and 10^26 for California's SB 53 and New York's RAISE Act.

Does frontier AI regulation apply to organisations that use the models? Mostly not. The frontier laws apply to developers. Deployers are covered by other rules, such as the EU AI Act's high-risk and transparency duties and UK data protection law.

Where it sits in the register

Third-party model and dataset provenance and inadequate AI vendor due diligence are the deployer's register entries; third-party model and vendor due diligence is where a developer's framework and system card become evidence.