EU AI Act GPAI rules are now enforceable - what actually changes

August 2 marked the date GPAI rules became legally binding. The framework exists; the enforcement capacity does not - yet.

In plain terms: The EU AI Act's rules for general-purpose AI models are now legally enforceable as of August 2, 2026. Labs must comply with transparency, documentation, and systemic-risk obligations—or face fines. Meta has not signed the voluntary code of practice; most others have.

The story: The transition from "agreed framework" to "binding law" is significant mostly for what it clarifies: the European AI Office is now the enforcement body, labs must maintain technical documentation and publish summaries of training data, and models above the 10^25 FLOP threshold face additional systemic-risk obligations, including red-team evaluations.

The practical gap remains large. The AI Office is understaffed relative to the scope of its mandate, the code of practice remains voluntary for now, and the first formal enforcement actions are unlikely before 2027. But the clock is running.

Meta's refusal to sign the voluntary code is a calculated bet: Llama's open-weight distribution complicates obligations that assume a centralized deployer. The EU's response to this structural ambiguity will set the tone for open-model regulation globally.

Under the hood: Systemic-risk designation (10^25 FLOPs training threshold) currently covers GPT-4-class models and above. Claude Opus-class, Gemini Ultra, and Llama 3-class models all fall under this regime.

So what: Compliance officers at AI companies serving European markets now have hard deadlines. The first real test will be when the AI Office issues its first formal inquiry—likely targeting a systemic-risk model with incomplete documentation.