Frontier AI

Also known as: frontier model

On this page 8

Frontier AI is the class of highly capable general-purpose AI models whose abilities match or exceed those of the most advanced systems available at a given time. In artificial intelligence, and especially in discussions of AI fundamentals and governance, the label marks the leading edge of capability, where new skills appear first and where risks are least understood. A single system at that edge is often called a frontier model.

The term is deliberately relative rather than technical. A model that counted as frontier in one year becomes ordinary once newer systems surpass it, so the boundary moves with the field. What stays constant is the combination of very large training compute, broad general-purpose capability and the possibility of dangerous capabilities that were not explicitly designed.

How frontier AI is defined in practice

Frontier AI is defined in practice by three overlapping criteria: capability relative to the state of the art, the scale of resources used to build the model, and its generality across tasks. No single criterion is sufficient, which is why policy documents usually combine a measurable proxy with a qualitative description.

The most common measurable proxy is training compute, counted in floating-point operations (FLOP). Compute is easy to verify before deployment and correlates with capability under empirical scaling laws, but it is an imperfect signal: algorithmic improvements let smaller training runs reach capabilities that previously required far more compute.

The table below summarizes the criteria that regulators, laboratories and researchers use to decide whether a model belongs to the frontier.

CriterionWhat it measuresMain weakness
Training computeTotal FLOP spent on trainingIgnores algorithmic efficiency gains
Benchmark performanceScores against the best existing modelsBenchmarks saturate and leak into training data
GeneralityBreadth of tasks without retrainingHard to quantify consistently
Dangerous capability evaluationsUplift in cyber, biological or autonomy tasksExpensive and only partly standardized
Development costHardware, energy and staff investmentNot disclosed by most developers
Criteria used to identify frontier AI

Compute thresholds and regulatory definitions

Regulators translate the idea of frontier AI into legal thresholds, most often expressed in training compute. The EU AI Act presumes that a general-purpose AI model has systemic risk when its cumulative training compute exceeds 10^25 FLOP, and it attaches obligations such as model evaluation, adversarial testing, incident reporting and cybersecurity protection.

Other frameworks have used a higher line of 10^26 FLOP, including a 2023 United States executive order that was later revoked and California legislation on frontier model transparency. These numbers are administrative triggers, not scientific boundaries, and most frameworks allow the authority to designate models below the threshold or to update it as efficiency improves.

FrameworkThreshold or testTypical obligation
EU AI ActAbove 10^25 FLOP or designationEvaluation, risk mitigation, incident reporting
US executive order of 2023Above 10^26 FLOP (revoked in 2025)Reporting of training runs and test results
California frontier model lawAbove 10^26 FLOP plus revenue criteriaPublished safety framework and transparency reports
Developer safety policiesCapability thresholds in evaluationsStronger safeguards before deployment
Selected approaches to defining frontier models

Because thresholds are fixed while hardware and algorithms improve, compute-based definitions drift over time. A model trained with ten times less compute may later reach the capability that once required the full threshold, which is why capability evaluations increasingly complement compute counts.

Frontier AI overlaps with several neighboring terms but is not synonymous with any of them. The distinction matters because policy, procurement and research discussions often use these labels interchangeably, which leads to confused obligations and misplaced expectations.

ConceptCore ideaRelation to frontier AI
Foundation modelBroad model adapted to many downstream tasksFrontier models are the most capable foundation models
Narrow AISystem built for one task or domainFrontier AI is general-purpose by definition
Artificial general intelligenceHypothetical human-level general abilityFrontier AI exists today and need not be AGI
Artificial superintelligenceHypothetical ability far beyond humansA speculative future endpoint, not a current category
Proprietary AI modelWeights kept private by the developerFrontier models may be proprietary or open-weight
How frontier AI differs from neighboring terms

The most frequent error is treating frontier AI as a synonym for AGI. Frontier status describes relative position in the present landscape, whereas artificial general intelligence describes an absolute level of capability that no system has been shown to reach.

Why frontier models raise distinct risks

Frontier models raise distinct risks because their capabilities are broad, partly unpredictable and widely accessible once deployed. Abilities such as multistep planning, code generation or detailed scientific reasoning can appear at scale without being specifically trained, a pattern often described as emergent or unanticipated capabilities.

Safety researchers group the main concerns into misuse, accidents and loss of control, and structural effects on markets and institutions. The list below covers the categories that frontier safety frameworks evaluate most often.

  • Cyber offense uplift: help with finding and exploiting software vulnerabilities.
  • Biological and chemical uplift: guidance that lowers barriers to producing dangerous agents.
  • Autonomous replication: ability of an agent to acquire resources and persist without oversight.
  • Persuasion and manipulation: large-scale generation of tailored deceptive content.
  • Concentration of power: dependence of many sectors on a few model providers.

Evaluating these risks is difficult because absence of evidence is not evidence of absence. A model may hide a capability under one prompting method and reveal it under another, so evaluators use red teaming, elicitation techniques and staged access rather than a single benchmark run.

How developers and governments manage frontier AI

Frontier AI is managed through a layered set of safeguards applied across the model lifecycle, from training decisions to post-deployment monitoring. Leading developers publish frontier safety frameworks that link specific capability thresholds to required protections, and public bodies such as the UK AI Security Institute run independent pre-deployment testing.

Lifecycle stageTypical controlPurpose
Before trainingCompute reporting and risk assessmentAnticipate capabilities early
During trainingInterim capability evaluationsDetect thresholds as they are crossed
Before releaseRed teaming and third-party testingFind misuse paths and failure modes
At releaseStaged access and usage policiesLimit exposure while risks are measured
After releaseMonitoring and incident reportingRespond to real-world harm quickly
Lifecycle controls for frontier models

A recurring debate concerns open-weight frontier models. Publishing weights supports research, auditing and competition, but it makes safeguards impossible to withdraw once released, so many frameworks require stronger evaluation before weights are made public.

Common misconceptions about frontier AI

Several misconceptions distort public and organizational decisions about frontier AI. Correcting them helps readers judge both the value and the limits of the most capable systems.

  • Bigger always means safer or better: scale improves average performance but does not remove hallucination or bias.
  • Frontier status is permanent: today’s frontier model becomes a mainstream model within a few years.
  • Compute alone defines the frontier: efficient architectures and better data can reach similar capability with less compute.
  • Only laboratories bear responsibility: deployers who fine-tune or integrate frontier models also carry obligations.

In plain words

Frontier AI means the most powerful AI systems that exist right now. Think of it like the fastest cars on a racing circuit: they set the pace, they are the most expensive to build, and they are the ones safety officials watch most closely.

Because these systems can do many different things, from writing software to explaining science, they can also be misused in ways that smaller tools cannot. That is why governments and developers test them more carefully. The key point is that frontier is a moving label: today’s leader becomes tomorrow’s standard model.

A real-world example

A mid-sized hospital network wants to use a general-purpose language model to summarize patient records and draft discharge letters. Its procurement team compares a frontier model offered through a commercial interface with a smaller open-weight model that can run on its own servers.

The frontier model produces clearer summaries and handles rare conditions better, but it falls under the provider’s systemic-risk obligations and requires data processing outside the hospital. The team runs a four-week pilot on 2,000 anonymized records, measuring factual errors and reviewer time. The frontier model cuts drafting time by roughly half, while the smaller model makes noticeably more omissions.

The outcome is a hybrid deployment: the frontier model drafts letters under a contract that excludes training on hospital data, and clinicians review every document before signing. The decision rests on measured performance and contractual safeguards, not on the frontier label alone.