Menu
Concepts

Frontier models

3 min readintermediateUpdated 28 Sept 2026
1 · In one line

Frontier models are highly capable general-purpose AI models at or beyond the capabilities of the most advanced current models.

1 · What it is

Frontier models sit near the moving edge of general-purpose AI capability. Frontier AI refers to the most advanced and capable systems at a given time.

Frontier safety work sets thresholds where unmitigated severe risks would be intolerable. A model is tested in defined risk categories, and its results are compared with thresholds. Capability thresholds can map to required security and deployment mitigations.

Risk assessment also considers model capabilities and the development and deployment context. UK DSIT guidance calls evaluation throughout the AI lifecycle vital. NIST’s misuse guidance covers risk management across that lifecycle.

2 · Why it exists

Model evaluations can inform decisions about training, security and deployment.

Moving boundaryFrontier AI refers to the most advanced and capable systems at a given time.
Severe risksSafety frameworks assess whether advanced capabilities could create severe harm.
Threshold responseCapability thresholds can map to required security and deployment mitigations.
3 · How it works

Follow one advanced model through a capability gate.

Capabilities Reports assess threshold crossings, while Safeguards Reports guide deployment decisions.
  1. 1 · evaluateTest the model for capabilities tied to defined risk categories.
  2. 2 · compareCompare evaluation results with explicit capability or risk thresholds.
  3. 3 · mitigateApply required security and deployment mitigations when a threshold is reached.
  4. 4 · decideUse capability and safeguards reports to inform deployment decisions.

Frontier AI refers to the most advanced and capable systems at a given time.

4 · Where it's used
WhoWhat they askWhat it works with
Evaluation team“Is this model approaching a severe-risk threshold?”Capability evaluation results
Security team“Do model weights need stronger protection?”Required security mitigations
Release committee“Are safeguards sufficient for deployment?”Capability and safeguards reports
5 · What it solves, and what it doesn't
solves
  • The label identifies highly capable general-purpose models near the current capability frontier.
  • Threshold frameworks connect measured capabilities to stronger safeguards.
  • UK DSIT guidance calls evaluation throughout the AI lifecycle vital.
doesn't solve
  • The frontier is not fixed; it refers to the most advanced systems at a given time.
  • Risk assessment also considers model capabilities and the development and deployment context.
  • The science for frontier safety continues to evolve.
6 · Go deeper
Foundation modelsRed teamingEvalsAI safetyAlignment

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. officialEmerging processes for frontier AI safety, UK Department for Science, Innovation and Technology · read 28 Sept 2026
  2. officialFrontier AI Safety Commitments, AI Seoul Summit 2024, UK and Republic of Korea governments · read 28 Sept 2026
  3. officialOur updated Preparedness Framework, OpenAI · read 28 Sept 2026
  4. officialAnthropic Transparency Hub - Voluntary Commitments, Anthropic · read 28 Sept 2026
  5. officialUpdated Guidelines for Managing Misuse Risk for Dual-Use Foundation Models, NIST · read 28 Sept 2026
  6. officialFrontier AI, UK National Cyber Security Centre · read 28 Sept 2026
Suggest a correction