Brussels set to name first systemic risk AI models this week under EU AI Act

The European Commission’s AI Office is expected to publish the first list of general purpose AI models designated as posing systemic risk as early as this week in Brussels, according to the Commission’s published application timeline and industry briefings. The move would mark the start of the EU AI Act’s most far reaching obligations for model providers, with potential fines up to 7 percent of global turnover for serious violations, as set out in the law.

Timing matters for enterprise users in Europe. Once named, affected providers face new documentation, safety testing, and incident reporting requirements that can shape product roadmaps and feature availability for the rest of 2026, according to the text of the Act and guidance from the AI Office.

The Act defines systemic risk general purpose models largely by a compute threshold of at least 10^25 floating point operations in training, along with other capability based criteria, according to the final compromise text approved by EU lawmakers in 2024. Thierry Breton, the EU commissioner for the internal market, has said the largest models will face the strictest rules, stating in 2024 that “no one is above the law” as the Commission stood up the AI Office.

Providers expected to be captured include OpenAI’s GPT 4 class and successors, Google’s Gemini Ultra tier, Anthropic’s Claude 3 family at the top end, Meta’s largest Llama generation, and Mistral’s flagship model, based on public disclosures of model scale and performance. Similarweb estimates showed ChatGPT accounted for a majority of consumer visits to generative AI sites in 2024, highlighting the reach of models likely to be scrutinized.

Designation triggers obligations for robust red teaming, reporting of serious incidents, disclosure of training compute and energy use, and measures to address systemic risks such as model proliferation and misuse, according to the Act and Commission briefings. The AI Office can require providers to undergo independent evaluations and to implement security controls proportionate to assessed risk.

Companies have been preparing. Anthropic published an updated Responsible Scaling Policy in 2024 that describes staged safety thresholds, Google DeepMind detailed its red team and evaluation work in model cards, and OpenAI has released system cards and abuse mitigation plans for major releases. Mistral, which raised about 600 million euros in mid 2024 at a reported 5.8 billion euro valuation, has argued for proportionate obligations on open models, a position echoed by parts of the French government.

What to watch this week

• Whether the first list is limited to a small set of frontier models or sweeps in widely used open models.
• Any immediate information requests or timelines set by the AI Office for compliance plans.
• Provider statements on region specific feature changes for EU users.

Why it matters for enterprises CIOs and product leaders should expect updated model terms, more detailed system cards, and possibly phased feature rollouts in the EU as providers align with the new regime. Coordination with NIST’s AI Risk Management Framework in the United States and the UK AI Safety Institute’s evaluation work remains a priority for multinationals that want consistent assurance across markets, since regulators on both sides of the Atlantic are watching the same set of large providers.