Delays to Frontier AI in the EU and UK
Analysis of 375 LLM Releases by Meta, Google, OpenAI, and Anthropic
This work represents the views of its authors, rather than the views of the organization, and does not constitute legal advice. GovAI technical reports have received extensive feedback, but have not gone through formal peer review.
- In recent years, frontier AI companies have sometimes delayed the release of new models to the EU and UK – or not released them at all. To what extent has this happened, and to what extent were regulatory barriers the cause?
- To investigate this, we collated and analyzed a dataset of 375 LLM releases by Meta, Google, OpenAI, and Anthropic over an eight-year period (June 2018 – May 2026), and used public sources to assess the most likely reason for delays or non-releases.
- By doing so, we hope to inform the design and implementation of AI regulation – in particular, how decision makers weigh regulatory objectives against the risk of models being delayed or withheld from their markets.
How often are models delayed or not released on the EU and UK markets?
- Relative to the US, we find that 11% of model releases were delayed or not released to the EU, and 7% were delayed or not released to the UK.
- The UK saw no delays in accessing the model ranked highest on the Epoch Capabilities Index at any point. The EU faced a 71-day delay for the web-app release of Claude 3 Opus while it ranked as the highest-performing model; API access, which developers use to build applications, was simultaneous with the US.
- Meta had the highest overall rate of delays and non-releases of any company in our dataset, with ~26% of its releases delayed or withheld in the EU and ~15% in the UK.
- We generally see a reduction in delays and non-releases over time. This could reflect companies transitioning from research to commercial organizations, with mature compliance functions and greater incentive to deploy broadly. It may also be due to greater regulatory certainty as EU and UK regulators clarified how data protection regulation applies to LLM-based services.
- In the first five months of 2026 there were no instances of a publicly released model being delayed or withheld from the EU and UK, but there were several instances of models being released to select entities under trusted-access programs (GPT 5.5 Cyber, Claude Mythos, GPT Rosalind). US export controls resulted in Anthropic pulling both the widely available Fable and the trusted-access version of Mythos. In the current environment, US national security concerns – rather than regulation – may pose the greater barrier to European access to frontier AI.
What are the causes of delays and non-releases?
- Regulatory barriers appear to be the primary cause of most delays or non-releases. Of the 68 examples of delays and non-releases in the dataset, we tentatively attribute regulatory factors as the primary, though not necessarily sole, cause of 56.
- Data protection regulation appears to be the main regulatory barrier, especially in relation to training on personal data. Non-text modalities (image, audio, real-time video) seem to face greater barriers than text.
- Although the UK and EU share broadly similar data protection laws, regulatory barriers appear greater in the EU for the period covered by our dataset. This is likely due to the EU's more aggressive regulatory enforcement and slower clarification of how data protection rules apply to the training and deployment of LLMs.
- We find no strong evidence of other regulations, including the EU AI Act, causing delays or non-releases. However, the AI Act's general-purpose AI provisions only came into force in August 2025, covering ten months of our dataset, and will only become enforceable in August 2026 (not covered at all by our dataset).
- In nine instances, the primary cause of a delay or non-release appears to be due to non-regulatory factors, including compute constraints, language support limitations, and broader product readiness concerns.
