Skip to content

EU — Frontier Models and GPAI

Relevant rules

Section titled “Relevant rules”
RuleProvisionsRelationship to GPAI
EU AI ActArts. 51–56; Annexes XI / XII / XIIIDedicated GPAI chapter
GPAI Code of PracticeAI Act art. 56Finalised Jul 10, 2025; signature = presumption of conformity
Digital Omnibus Proposal—Nov 2025 proposal to postpone high-risk provisions to Dec 2027 (does not affect GPAI)

Structure of the GPAI chapter (arts. 51–56)

Section titled “Structure of the GPAI chapter (arts. 51–56)”

Two-tier classification (art. 51)

Section titled “Two-tier classification (art. 51)”

All GPAI models (open-source or closed-source):

  • Defined as: general-purpose, large-scale self-supervised trained, capable of performing diverse downstream tasks.

GPAI models with systemic risk:

  • Automatic presumption: cumulative training compute ≥ 10²⁵ FLOP.
  • Commission designation: based on parameter count, dataset size, user count, and capability benchmarks.
  • Operators may apply for re-assessment.

Obligations for all GPAI (art. 53)

Section titled “Obligations for all GPAI (art. 53)”
  1. Technical documentation (Annex XI): training process, evaluation results, capability description.
  2. Downstream documentation (Annex XII): for downstream providers.
  3. Copyright policy: establish a policy to comply with the EU Copyright Directive’s TDM exception.
  4. Public summary of training data: template issued by AI Office (published Jul 2024).

Additional obligations for systemic-risk GPAI (art. 55)

Section titled “Additional obligations for systemic-risk GPAI (art. 55)”
  • Model evaluation: including adversarial testing and red-teaming.
  • Systemic-risk assessment and mitigation.
  • Serious-incident reporting (art. 56) → AI Office.
  • Cybersecurity: model plus physical infrastructure.

Code of Practice

Section titled “Code of Practice”

See the dedicated rule page GPAI Code of Practice. Key points:

  • Finalised Jul 10, 2025 (roughly 1,000 participants across multiple drafting rounds).
  • Aug 1, 2025: Commission + AI Board issue Adequacy Decisions.
  • Three-chapter structure: Transparency, Copyright, Safety and Security.
  • Signature = presumption of conformity (art. 56(8)).
  • Signed (as of Apr 2026): Anthropic, Google DeepMind, Microsoft, Mistral (full); OpenAI (with reservations); Meta (dissent on the copyright chapter); xAI (dissent on the safety chapter).
  • Not signed: DeepSeek, ByteDance, Zhipu, and other Chinese firms.

Open-source exemption (art. 53(2))

Section titled “Open-source exemption (art. 53(2))”

Open-source GPAI is exempt from some obligations (technical documentation, downstream documentation), subject to:

  • A “free and open-source licence”.
  • Weights, architecture, and usage methods are public.
  • A training-data summary must still be published.
  • Systemic-risk GPAI is not exempt even if open-sourced.

Effect: Meta LLaMA-like open series benefit, but only partially. Open-source models like DeepSeek’s are subject to the same rules if they enter the EU market.

10²⁵ FLOP presumption in practice

Section titled “10²⁵ FLOP presumption in practice”

Models presumed above the threshold as of Apr 2026 (estimated):

  • GPT-4 / 4o / o1 / o3 series (OpenAI)
  • Claude 3 / 4 series (Anthropic)
  • Gemini 1.5 / 2 Ultra (Google DeepMind)
  • Some national-scale research models (French Mistral Large 2 may qualify).

The presumption is rebuttable: a provider may argue that a specific model does not rise to systemic risk, but must submit evidence to the AI Office.

Timeline

Section titled “Timeline”
DateEvent
Aug 1, 2024AI Act in force
Jul 10, 2025Code of Practice finalised
Aug 1, 2025Commission + AI Board issue Adequacy Decisions; signatory list published
Aug 2, 2025GPAI obligations (arts. 51–56) applicable
Aug 2, 2026Remaining provisions applicable
Aug 2, 2027Specific embedded high-risk cases

Comparison with China and the US

Section titled “Comparison with China and the US”
  • Predictability: the EU leads (explicit compute threshold + documentation checklist + code of conduct).
  • Binding force: the EU is strongest (hard law + penalties of 3% of global annual turnover).
  • International reach: the Brussels Effect is in motion. Documentation prepared by model providers for the EU market effectively becomes the disclosure baseline in other jurisdictions.