Examples
Worked examples
- Is an instance
A model whose training run exceeded 10^25 FLOPs and that is subject to systemic-risk obligations under the EU AI Act.
- Is an instance
A frontier-lab model voluntarily disclosed to the UK AISI under the pre-deployment access agreement.
Counter-examples
Looks similar, but isn't
- Not an instance
A 7B-parameter open-weight model usable on a single GPU.
- Not an instance
A task-specific classifier trained for one application.
Editorial commentary
Frontier model is a relative, policy-oriented term for a model whose capabilities meet or exceed the most advanced publicly known systems at a given point in time — as opposed to foundation model, which describes a training paradigm and does not expire as newer systems are released. The Frontier Model Forum, formed in 2023 by several major AI developers, defines the category around capability — highly capable foundation models that could exhibit dangerous capabilities — with no fixed compute number attached to the definition.
The genuinely contested part: how “frontier” gets operationalised
Regulators need a bright line that an industry capability judgement does not provide, so several jurisdictions have attached a training-compute threshold to the concept instead. Under the EU AI Act, a general-purpose AI model is presumed to carry systemic risk — triggering the Act’s heaviest obligations (model evaluation, adversarial testing, incident reporting) — once cumulative training compute exceeds 10^25 floating-point operations (Article 51(2)); this presumption is rebuttable, and the European Commission can adjust the threshold by delegated act as capability at a given compute level rises. As of April 2026, roughly a dozen released models were reported to exceed that threshold. A separate, since-lapsed US mechanism, Executive Order 14110 (October 2023), had set a 10^26-FLOP reporting threshold for dual-use foundation models; it was revoked in January 2025 by Executive Order 14179, which replaced it with a deregulatory approach carrying no equivalent FLOPs threshold. As of this writing, the EU AI Act’s 10^25-FLOP presumption is the only FLOPs-based regulatory line still active among major jurisdictions, and any source citing a current “US federal FLOPs threshold” should be treated as out of date.
References
- EU AI Act, Article 51(2) (artificialintelligenceact.eu).
- Executive Order 14179, “Removing Barriers to American Leadership in Artificial Intelligence” (Jan. 2025), revoking Executive Order 14110 (Oct. 2023).
- See also: Compute (FLOPs estimate), EU AI Act, AI safety case.
Also known as
frontier AI model
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Frontier model"
vocab-term-identifier="https://casrai.org/dictionary/term/frontier-model" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/frontier-model",
"name": "Frontier model",
"identifier": "https://casrai.org/dictionary/term/frontier-model",
"description": "A foundation model whose capabilities meet or exceed the most advanced publicly known systems at the time of training, often defined operationally by training-compute thresholds or by performance on canonical benchmarks.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/frontier-model",
"alternateName": [
"frontier AI model"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-09-05T14:24:41",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
