Foundation models are defined by how they are trained and reused: they learn from broad data at scale and can be adapted to many tasks. Frontier models are defined by their position at the leading edge of capability—or, in some safety-policy work, by whether they may have dangerous capabilities. The labels are not opposites: a model can be both, and not every foundation model is frontier.
What is a foundation model?
Stanford’s Center for Research on Foundation Models (CRFM) describes foundation models as models trained on broad data at scale that can be adapted to a wide range of downstream tasks. The term describes a reusable base, not necessarily a finished system for a particular job. A model may need task-specific adaptation before it is useful in a given application. See Stanford CRFM’s 2021 report, On the Opportunities and Risks of Foundation Models.
What does “frontier model” mean?
“Frontier model” has no single universal definition in the sources cited here. It is used in at least two ways, so the surrounding context matters.
At the capability frontier
In a capability-relative sense, a frontier model is near or beyond the average capabilities of the most capable models then available, and may differ in scale, design, or the mix of capabilities and behaviors it shows. This use is relative to the state of the field: a model’s position can change as new models appear. Shevlane and coauthors describe this framing in their 2023 paper, Model evaluation for extreme risks.
#1 Best Overall
In safety-policy discussions
Some policy work uses “frontier AI model” more narrowly, connecting high capability to the possibility of sufficiently dangerous capabilities. Anderljung and coauthors state: “For the purposes of this paper, we define ‘frontier AI models’ as highly capable foundation models that could exhibit sufficiently dangerous capabilities.” The phrase “for the purposes of this paper” matters: this is a scoped definition, not a universal standard. Their 2023 paper is Frontier AI Regulation: Managing Emerging Risks to Public Safety.
How the two labels differ
| Question | Foundation model | Frontier model |
|---|---|---|
| What does the label describe? | Broad training and adaptability across downstream tasks. | Either a position near the leading edge of capability or, in a stated safety-policy definition, potential dangerous capabilities. |
| How is it identified? | Look for broad data, large-scale training, and transfer or adaptation to different tasks. | For capability-relative use, compare with the strongest existing models and consider scale, design, and capability mix. For policy use, examine dangerous capabilities and possible severity. |
| Is there a fixed boundary? | It is a broad technical concept; individual usage can vary. | No universal threshold is established by the definitions cited here. The criterion depends on context. |
| Can a model have both labels? | Yes. | Yes. Under the cited policy definition, frontier AI models are a subset of foundation models. |
Are frontier models the same as foundation models?
No. The terms answer different questions. “Foundation” concerns how a model is trained and whether it can be adapted across tasks. “Frontier” concerns where capability stands relative to the field, or whether a policy definition’s risk criterion is met. They are not competing architectures or product categories.
Rank #2
It follows that being a foundation model alone does not make a model frontier. Nor does being state of the art by itself establish that a model meets a dangerous-capability threshold. Capability comparisons and risk assessments are distinct; the latter require an explicit criterion rather than an assumption based on rank.
How to interpret the term in an article or policy
- Check how the author defines “frontier.” Is the claim about relative capability, or about potentially dangerous capability?
- Look for the comparison or risk criterion. A capability claim should be understood in relation to the strongest models at that time; a policy claim should specify the relevant dangerous capabilities and severity.
- Keep the date in view. A capability-relative label can become outdated when stronger models arrive. The 2023 capability framing is not a current ranking of models.
- Do not treat the labels as interchangeable. A foundation model may be widely adaptable without being at the capability frontier; a policy use of “frontier AI” adds a risk condition.
What the cited risk statistic does—and does not—say
Shevlane and coauthors’ 2023 paper reports that 36% of AI researchers surveyed in 2022 thought AI systems could plausibly cause a catastrophe this century at least as bad as an all-out nuclear war. The paper attributes the survey to Michael and coauthors (2022). This is a report of respondents’ views, not an estimate that such a catastrophe has a 36% probability.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
