EU keeps August 2026 AI deadline

Europe’s AI Act has moved out of the abstract and into the calendar. The big fight now is not whether rules for powerful foundation models exist, but when they bite. And on that point, the answer is clearer than many labs wanted: the EU’s general-purpose AI rules started applying on August 2, 2025, and enforcement for new models lands on August 2, 2026. Older models get one extra year, but that’s it. (digital-strategy.ec.europa.eu) ### What exactly is on the clock? This is about general-purpose AI models — the broad models that can be integrated into lots of downstream products, not just a single narrow tool. The Commission’s guidance says the obligations for providers of those models entered into application on August 2, 2025. The AI Office can enforce those rules one year later for new models, and two years later for models already on the market before that date. (digital-strategy.ec.europa.eu) ### What do providers actually have to do? For all general-purpose model providers, the core package is pretty concrete. They have to draw up technical documentation, put a copyright policy in place, and publish a summary of the training content. If a model is considered to pose systemic risk, the obligations get heavier — notification to the Commission, risk assessment and mitigation, serious-incident reporting, and cybersecurity protections. (digital-strategy.ec.europa.eu) ### Why does August 2026 matter so much? Because that is when the soft-preparation phase turns into a real enforcement phase for newly placed models. A lot of companies can live with vague future regulation. A dated enforcement trigger is different — it forces legal, policy, safety, and engineering teams to build repeatable compliance systems now. (digital-strategy.ec.europa.eu)es, and deciding who in Europe is legally on the hook. (digital-strategy.ec.europa.eu) ### What counts as a “new” versus “existing” model? The line the EU uses is whether the model was placed on the market before August 2, 2025. Models already on the market before that date have until August 2, 2027 to comply. Models placed on the market on or after August 2, 2025 fall into the earlier track, with enforcement from August 2, 2026. That sp(digital-strategy.ec.europa.eu)e old model or functionally a new one. (digital-strategy.ec.europa.eu) ### Is the Code of Practice the law? No — but it is the EU’s preferred compliance shortcut. The General-Purpose AI Code of Practice is voluntary, written to help providers show they meet the AI Act’s safety, transparency, and copyright obligations. The Commission and Member States have said providers that sign and follow it should get lower compliance burden and more legal certainty. Basically, the law is the law, but the code is the map the EU is handing companies. (digital-strategy.ec.europa.eu) ### Which models get the toughest treatment? The models presumed to pose systemic risk. The AI Act guidance uses a compute threshold of more than 10^25 FLOP for that presumption, though the threshold is under review. Those providers face the extra safety stack — evaluations, incident reporting, and cybersecurity — because the EU treats the most capable models less like ordinary software and more like infrastructure that can fail at scale. (digital-strategy.ec.europa.eu) ### So what’s the real takeaway? The window for hoping this would slide has basically closed. For major labs selling into Europe, 2026 is no longer a distant policy milestone. It is an operational deadline. The winners here are probably the companies that can turn compliance into a product discipline early, because once enforcement starts, “we’re still figuring it out” stops sounding like a strategy.