Summary As proposed, CADA's "frontier AI" and the EU AI Act's "general-purpose AI model" (GPAI) come at advanced AI from opposite directions. CADA Article 2(4) defines frontier AI by capability relative to the field — models (or systems built on them) that perform a wide variety of tasks and approach, reach, or exceed the current state of the art. The AI Act instead targets GPAI models by their generality and, for the strictest obligations, by systemic risk. CADA's label is a supply-side gateway to EU support; the AI Act's GPAI/systemic-risk regime is a compliance burden. The two can apply independently.

Detail

The proposed CADA and the existing EU AI Act classify advanced models on different bases. Telling CADA's "frontier AI" apart from the AI Act's "general-purpose AI model" matters for anyone planning around both.

CADA's definition of frontier AI

CADA introduces "frontier AI" as a category for strategic support. Article 2(4) defines it as:

"AI models or AI systems built upon such models that can perform a wide variety of tasks and that approach, reach or exceed the current state of the art"

This is a capability-relative test. It sets no parameter count, architecture, or compute figure. It asks how the model performs against the current state of the art — whether it is among the most capable globally across a wide range of tasks.

The term is operationalised through Article 8, under which the Commission may, by decision, recognise frontier AI priority projects against set criteria. Article 9 then supports the allocation of AI computing resources to those priority projects (alongside industrial and physical AI projects), with the Union and Member States providing sufficient compute. The orientation is scaling up strategic assets and reducing dependence on third-country technology.

The AI Act's general-purpose AI model

The EU AI Act (Regulation (EU) 2024/1689) uses "general-purpose AI model" (GPAI). Its focus is not whether a model is the best in the world but its generality and capacity to be integrated into many downstream systems, with a separate, stricter tier for GPAI models that present systemic risk. The AI Act's GPAI provisions — including the systemic-risk obligations and the Code of Practice — sit in its Articles 51 to 56.

Key difference: relative capability vs the AI Act's systemic-risk gate

The most consequential difference is what triggers action.

1. Moving target vs structural test. CADA's frontier AI is dynamic: as the state of the art advances, a model that qualified last year may not this year. The AI Act's GPAI test turns on the model's generality and its placement on the market, with systemic risk assessed for the high-impact tier. (The AI Act sets a training-compute threshold as one route into the systemic-risk presumption; for the exact figure and mechanics, consult the AI Act and its delegated/implementing measures rather than relying on a number here.)

2. Support vs compliance. CADA's frontier AI designation is a supply-side mechanism: recognition as a frontier AI priority project under Article 8 can unlock matched AI computing resources under Article 9. The AI Act's GPAI/systemic-risk regime is a demand-side compliance mechanism, imposing obligations on model providers.

3. Scope. CADA's definition reaches "AI models or AI systems built upon such models," explicitly extending to downstream systems built on a frontier model. The AI Act draws a clearer line between the model (GPAI) and downstream AI systems, applying different obligations to each.

Why the distinction matters for implementation

For a CTO, you may be operating under both frameworks at once:

  • If your model is among the most advanced globally, it may be frontier AI under CADA and a candidate for a frontier AI priority project (Article 8) eligible for EU compute support.
  • If your model meets the AI Act's GPAI criteria — and, for the strictest tier, presents systemic risk — you face the AI Act's GPAI obligations regardless of CADA status.

The two can diverge: a highly efficient model could be frontier under CADA without crossing the AI Act's systemic-risk gate, and a model could fall under the AI Act's GPAI/systemic-risk regime without being "frontier" if it does not approach the state of the art across a wide range of tasks.

What this means for you

For CTOs and architects:

  • Eligibility for support. If you build advanced models, assess them against CADA's frontier AI criteria (Article 2(4)). If they approach or exceed the state of the art, consider recognition as a frontier AI priority project (Article 8) for access to compute resources under Article 9.
  • Compliance planning. Independently, assess your AI Act exposure. If your model is a GPAI — particularly one in the systemic-risk tier — prepare for the corresponding AI Act obligations regardless of CADA status. Track your training compute and benchmark results, since these feed different determinations.
  • Documentation. Keep two distinct evidence bases: capability relative to the state of the art (for CADA) and the metrics relevant to the AI Act's GPAI/systemic-risk assessment.
  • Strategic positioning. CADA's frontier AI focus aligns with reducing dependence on non-European providers; positioning models as frontier AI can raise their value to public sector and research buyers seeking sovereign options.

Common misconceptions

  • Misconception: "Frontier AI" and "GPAI" mean the same thing.
    • Reality: They do not. Frontier AI (CADA) is about capability leadership; GPAI (AI Act) is about generality and, for the strict tier, systemic risk.
  • Misconception: Being frontier AI under CADA automatically triggers the AI Act's systemic-risk obligations.
    • Reality: No. The frameworks operate independently; a CADA frontier model need not fall into the AI Act's systemic-risk tier.
  • Misconception: CADA sets a fixed compute threshold for frontier AI.
    • Reality: CADA uses a relative capability test ("approach, reach or exceed the current state of the art"), not a compute figure — more flexible, but more subjective, requiring benchmarking against the field.

Official sources

Related

This is general information about a draft EU regulation, not legal advice.