Home / GPAI

General-purpose AI models

Model-level rules, in force since August 2025 — independent of, and stacking with, system-level risk classes. On August 2, 2026 the AI Office's enforcement powers begin, with an expanded supervisory scope under the Omnibus.

THE SYSTEMIC THRESHOLD
10²⁵ FLOP
training compute — or Commission designation. Triggers the Art. 55 tier.
ENFORCEMENT
AUG 2, 2026
AI Office powers begin · carve-outs: law enforcement, border, judicial, financial stay national.
LEGACY MODELS
AUG 2, 2027
models placed before Aug 2, 2025 must be brought into compliance.

Duties by tier

Art. 53(1)(a) ALL GPAI Technical model documentation Maintain documentation of training, testing and evaluation — available to the AI Office and national authorities on request. Art. 53(1)(b) ALL GPAI Downstream provider information Give system builders what they need to understand capabilities, limitations and comply with their own duties. Art. 53(1)(c) ALL GPAI Copyright policy A policy to comply with EU copyright law, including honouring text-and-data-mining opt-outs. Art. 53(1)(d) ALL GPAI Public training-data summary Publish a sufficiently detailed summary of training content, using the AI Office template. Art. 55(1)(a–b) SYSTEMIC Model evaluation & adversarial testing State-of-the-art evaluations including red-teaming; assess and mitigate systemic risks at Union level. Art. 55(1)(c–d) SYSTEMIC Incident reporting & cybersecurity Track and report serious incidents to the AI Office without undue delay; ensure adequate cybersecurity for model and infrastructure.
Source: Art. 53–55 · AI Office guidelines · verified 2026-07-16 · Open-weight models get partial exemptions unless systemic.

The two codes of practice

GPAI Code of Practice SIGNABLE · SIGNATORIES SINCE AUG 2025

The Commission-endorsed route to Art. 53/55 compliance: transparency, copyright, safety and security chapters. Signing simplifies your evidence burden; not signing means proving equivalence another way.

Signatory list & chapters →
Transparency Code (AI-generated content) FINAL · JUN 10, 2026

Marking and detection practices for providers; deepfake and public-interest-text labelling for deployers. The Commission and AI Board consider it an effective route to Art. 50 compliance — directly relevant to the Aug 2 deadline.

How it maps to Art. 50 →
Downstream note: if you fine-tune or build on a GPAI model, you may inherit provider duties for the resulting system — and the model provider owes you the documentation to comply. OBL-0060 →