Frontier AI

Working Definition

Frontier AI refers to general-purpose artificial intelligence models that match or exceed the capabilities of the most advanced systems existing at a given moment and may therefore possess dangerous capabilities posing severe risks to public safety. By design, the concept is relational and moving: it describes not a fixed architectural property, but a temporary position at the leading edge of the capability distribution—a status that today’s flagship model typically loses within twelve to eighteen months. In everyday industry and media usage, “frontier model” simply means the latest flagship release from one of a handful of leading laboratories; in regulatory and governance usage, it demarcates the narrow subset of foundation models subject to heightened oversight obligations such as pre-deployment red-teaming, compute reporting, and incident disclosure. Because capabilities cannot be measured before a model is trained, legal texts routinely operationalize the category through proxy thresholds of total training compute measured in floating-point operations (FLOP).

Genesis and Global Lineage

The term was coined in July 2023 by 24 researchers affiliated with the Centre for the Governance of AI (GovAI), OpenAI, and partner universities in Frontier AI Regulation: Managing Emerging Risks to Public Safety (Markus Anderljung, Joslyn Barnhart, Anton Korinek, Jade Leung, Cullen O’Keefe, Jess Whittlestone, et al.), defining “what we term ‘frontier AI’ models.” The coinage solved a two-sided regulatory dilemma: Stanford’s 2021 category of foundation models (Rishi Bommasani et al.) was too broad, sweeping in thousands of small academic and open-weight checkpoints posing no catastrophic risk, whereas artificial general intelligence (AGI) was too contested and futuristic for statutory law. Institutional uptake was immediate: on July 26, 2023—twenty days after the arXiv preprint—Anthropic, Google, Microsoft, and OpenAI formed the Frontier Model Forum; in September 2023 Anthropic released its Responsible Scaling Policy with AI Safety Levels modeled on biosafety labs; and in November 2023 the UK government made “frontier AI” the centerpiece of the Bletchley Park AI Safety Summit. Two translations quickly altered the concept: lawmakers in Washington and Brussels converted the qualitative definition into numerical compute thresholds (10²⁶ FLOP in the October 2023 U.S. Executive Order; 10²⁵ FLOP for general-purpose AI models with “systemic risk” in the 2024 EU AI Act), while corporate marketing teams co-opted “frontier” as a badge of prestige.

Central European Lineage

The only documented link to the region is modest and biographical, concerning the intellectual formation of one coauthor of the foundational paper. Anton Korinek—third among the 24 authors of Frontier AI Regulation (2023) and a leading economist studying the macroeconomic implications of transformative AI—received his foundational legal and economic education at the University of Vienna (earning his cand. iur. in 1997 and his Master’s degree in economics in 2000 under Alexandre Lamfalussy, before completing his doctorate at Columbia University). In the 2023 paper, Korinek’s contribution reflects the economic analysis of regulatory externalities rather than the coining of the term, and we record this connection transparently without overstating its weight. More broadly, no earlier Central European text formulated the regulatory concept of a moving capability frontier. Within the contemporary European AI ecosystem, regional sovereign-language initiatives—such as Poland’s open-weight language models Bielik and PLLuM—define their strategic niche explicitly by contrast with the American and Chinese “frontier,” optimizing for parameter efficiency, local legal compliance, and cultural-linguistic fidelity well below the 10²⁵ FLOP systemic-risk line.

Why Now

The concept moved from an arXiv preprint to international summit declarations in four months in 2023 because GPT-4 (March 2023) forced Western governments to decide whom to regulate without strangling the wider software economy. Neither Washington nor London wanted heavy licensing duties on startups fine-tuning seven-billion-parameter models, yet both wanted visibility into billion-dollar training runs that might assist with cyberattacks or biological weapons design. The UK’s Frontier AI Taskforce (renamed the AI Safety Institute) and the Bletchley Declaration (November 2023) installed the term in diplomatic vocabulary, while U.S. Executive Order 14110 (October 2023) tied mandatory red-teaming disclosures to 10²⁶ FLOP. Notably, the EU’s AI Act (2024) refused the word “frontier,” substituting “general-purpose AI models with systemic risk” (10²⁵ FLOP). Between 2024 and 2026, two developments strained the concept: first, algorithmic efficiency and test-time compute (highlighted by Chinese models like DeepSeek-R1 in early 2025) achieved frontier-level reasoning at a fraction of historical pretraining FLOP thresholds; second, the International AI Safety Report 2025 chaired by Yoshua Bengio deliberately used “general-purpose AI” throughout, marking a quiet retreat from a term entangled with industry lobbying.

Explanatory Power

The concept highlights three structural dynamics of contemporary AI policy. First, regulatory risk is treated as positional rather than intrinsic: the same capability that triggers government notification when first achieved at the frontier ceases to be “frontier AI” eighteen months later when distilled into a smaller open-weight model running on a laptop—exposing a built-in temporal decay in frontier-only regulation. Second, it shows how legislation targets five or six corporations without naming them: setting a threshold at 10²⁵ or 10²⁶ FLOP isolates the handful of hyperscalers and labs spending hundreds of millions of dollars on a single training run, clarifying why those firms initially welcomed frontier governance and why smaller competitors denounced it as regulatory capture. Third, it illuminates the growing evaluation debt of AI safety: because a frontier model exceeds all existing systems, every generation requires inventing novel biosecurity, cyber-offense, and autonomy evaluations after training. This explains why battles over California’s vetoed SB 1047 (2024) revolved around a $100 million compute threshold, and why labs publish hundred-page system cards centered on CBRN (chemical, biological, radiological, and nuclear) uplift tests.

Axes of Debate

Underappreciated Angle

The word frontier does far more ideological work than its technical definitions admit. In the American political imagination—crystallized by historian Frederick Jackson Turner’s 1893 lecture The Significance of the Frontier in American History—the frontier is not a hazardous industrial perimeter, but a redemptive space of manifest destiny where pioneers forge ahead, law arrives late, and halting expansion means stagnation. Imported into AI governance, that metaphor pre-decides the normative frame: pushing the boundary outward becomes a heroic historical vector, and regulation is reduced to a supply wagon managing collateral risk. Contrast this with the EU’s statutory phrase “general-purpose AI models with systemic risk”—borrowed from post-2008 banking supervision, where a “systemically important institution” is not a romantic pioneer on the prairie, but an infrastructure concentration too big to fail cleanly. By choosing the word frontier, global tech policy adopted a vocabulary that romanticizes the capability race it claims to restrain.