EU Commission publishes GPAI implementing act — frontier-model thresholds set, capability-eval reporting requirements active August 1
The European Commission published the long-pending GPAI implementing act this week, fixing the systemic-risk threshold for general-purpose AI models at 10^26 training FLOPs and establishing the capability-evaluation reporting framework that goes active August 1. The threshold value is materially below the 10^25 FLOPs in the original Codes of Practice draft and captures all current frontier models.
The 10^26 FLOPs threshold value is calibrated to include GPT-5, Claude 4.6/4.7 Opus, Gemini 2.5 Pro, and Llama 5.x 405B. Below the threshold, providers face standard transparency and copyright-compliance requirements. Above the threshold, providers face capability-evaluation reporting, systemic-risk assessment, and incident-disclosure obligations.
The August 1 active date gives frontier providers ten weeks to set up the reporting infrastructure for the capability-evaluation framework. The framework cross-references the UK AISI Methodology 2.0 protocol, which means EU GPAI compliance now requires the activation-probe-testing infrastructure that only Anthropic, Google DeepMind, and Apollo Research currently operate end-to-end.
EurActiv — GPAI implementing act published → · Politico EU — EU sets 10^26 FLOPs systemic-risk threshold → · EurActive AI Watch — What the August 1 deadline does to frontier providers →