EU begins AI Act enforcement: RFIs to OpenAI, Anthropic, Google
First requests under the GPAI rules put training-data summaries and safety evals on the clock — €15M-or-3% fines, EU access at stake. Plus a 770B open model.
Copy markdown
The RFIs, and what they demand
The EU's AI Office issued its first formal Requests for Information to OpenAI, Anthropic, and Google, asking for training-content summaries and details on how frontier models are evaluated and monitored. It's the first hard enforcement step since the general-purpose AI obligations took effect Aug 2.
The teeth: €15M or 3% of revenue
Refusing, stalling, or sending misleading answers can draw fines up to €15M or 3% of global revenue, and ultimately restrictions on offering the model in the EU. The rules reach any provider serving a general-purpose model to EU users, wherever the company is based.
What it means if you ship to the EU
Expect model cards, terms, and training-data disclosures to shift as providers answer. If your product serves EU users, watch each provider's EU availability notes — this is the 'who gets what' lever, and it's now live rather than theoretical.
Elsewhere: Tencent open-sources Hy4-Preview
Tencent released the weights for Hy4-Preview, a 770B-parameter MoE (49B active) with a 1M-token context, live on OpenRouter and Tencent Cloud at ~$0.83/M input and ~$2.50/M output. Tencent's own eval edges GLM-5.3 and Kimi K3 on engineering tasks — a cheap long-context open model to route coding and agent work to today.