Markus Anderljung, GovAI policy director, finds 11% of frontier AI models are delayed or withheld in the EU

In the first five months of 2026, nothing was held back from Europe due to domestic regulation.

However, what did keep frontier AI out was US export controls – the order to pull Anthropic's Fable and Mythos.

Going forward, US national security concerns seem likely to play a bigger role than domestic regulatory barriers.

Delays have also fallen over time, as regulators clarified how data-protection rules apply.

We find no strong evidence the EU AI Act has caused delays – but its GPAI rules only become enforceable in August 2026, after our data ends.

How often do EU and UK users actually miss out on frontier AI because of regulation? We went through 375 LLM releases from Meta, Google, OpenAI, and Anthropic since 2018. 11% were delayed or withheld in the EU, 7% in the UK. Less than the discourse suggests.

And access to the best models has *mostly* been fine. The UK never faced a delay on the model ranked highest on Epoch's Capabilities Index. The EU's worst case was a 71-day delay to the Claude 3 Opus web app – though API access was simultaneous with the US.

When there are delays of models, they tend to be longer in the UK and then EU. In the EU, the median was 53 days' delay and in the UK, 15 days.

OpenAI had frequent short delays, whereas Meta had a smaller number of longer delays.

Full report, with the data and method – led by John Lidiard and Oleksandra Vereschak, with Tom Gibbs and me at @GovAIOrg: https://www.governance.ai/research-paper/delays-to-frontier-ai-in-the-eu-and-uk

Dataset: https://docs.google.com/spreadsheets/d/1mAaJyAiVxg6y-_XbJp3I8sAqrckSdS79I1b-VbQvjvw/edit?gid=930881061#gid=930881061

Huge thanks to the team who put this data together!