EU AI Act GPAI Rules Go Live

What this countdown tracks

Sunday 2 August 2026 — the application date for the EU AI Act's General-Purpose AI (GPAI) provisions under Article 113. From this date the obligations on providers of general-purpose AI models — including model documentation, training-data summaries, copyright-policy disclosures and systemic-risk duties for the most capable models — become directly enforceable across the European Union, with administrative fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.

About the AI Act

Regulation (EU) 2024/1689 — the AI Act — entered into force on 1 August 2024 with a staggered application schedule. Prohibitions on unacceptable-risk AI systems took effect on 2 February 2025. The GPAI chapter (Articles 51–55 and Annex XI–XIII) takes effect on 2 August 2026, alongside governance, penalties and member-state authority designations. High-risk AI systems under Annex III follow on 2 August 2027.

The GPAI obligations apply to providers of any model trained on more than 10²³ FLOPs of compute (the basic GPAI threshold) and impose heightened systemic-risk duties on models above 10²⁵ FLOPs — a threshold that captures GPT-4-class and larger models including the publicly disclosed leading models from OpenAI, Anthropic, Google DeepMind, Meta, xAI and Mistral.

Key obligations

• Maintain up-to-date technical documentation under Annex XI.
• Make available a detailed summary of training data to the EU AI Office.
• Adopt a copyright-compliance policy, including respect for opt-outs under the DSM Directive.
• For systemic-risk models: model evaluations, adversarial testing, serious-incident reporting, cybersecurity safeguards, energy-efficiency disclosures.
• Designate an EU authorised representative if not established in the Union.

Stakes

• First enforceable AI-specific regulation against frontier models anywhere in the world.
• Penalties of up to €15M or 3% global revenue — a credible threat against trillion-dollar AI labs.
• The Code of Practice for GPAI — a voluntary instrument that demonstrates compliance — was published by the AI Office in mid-2025; signatories include most major labs.
• Brussels-effect pressure: many providers will apply EU-aligned documentation worldwide.

Significance

2 August 2026 is the date the EU AI Act becomes a regulatory reality for frontier-model providers, with the AI Office in Brussels exercising direct supervisory authority. National competent authorities under each member state's designation also begin parallel oversight of downstream deployers.

How to follow

The European Commission's AI Office (digital-strategy.ec.europa.eu/en/policies/ai-office) is the primary enforcement body. Member-state designations are published in the Official Journal. Civil-society watchers include EDRi, the Future of Life Institute and Access Now; industry trackers include the Centre for European Policy Studies, Bruegel and the AI Now Institute.

Related countdowns

Pair with the existing EU AI Act enforcement 2026 and EU AI Act phase 2 2027 milestones, plus the NYT v. OpenAI verdict, the Musk v. Altman trial and the GPT-6 launch window.

FAQ

When do GPAI rules apply? 2 August 2026. What are the penalties? Up to €15 million or 3% of global turnover, whichever is higher. Who enforces? The EU AI Office in Brussels plus national competent authorities. Does it apply to US-based labs? Yes, if their models are placed on the EU market.