Open to Interpretation: How Frontier AI Developers Practice Openness through Licensing

Openness is a central but contested category in global artificial intelligence (AI) governance, yet policy debates often treat it as a binary property of developers, or even of states. This article asks how leading foundation model developers practice openness through release and licensing decisions, and what this implies for how openness is governed globally. We analyze an original longitudinal dataset of 1,313 model releases by twelve developers in the United States (US), China, France and Canada between January 2023 and October 2026. We find that openness is a property of model portfolios rather than of developers. We distinguish four configurations, defined by whether, and on what terms, a developer releases its most capable model: permissive, conditional, or non-commercial open flagship, or closed flagship with permissively licensed secondary models. Secondary models have converged on generic permissive licenses, while open flagships have moved towards bespoke conditions that target revenue rather than users. The sharpest division concerns whether third parties can inspect a developer’s most capable model without permission: all US developers now withhold this, while developers elsewhere, except Baidu, allow it. Drawing on the notion of a "polysemous compromise” and on practice-based accounts of norm formation, we argue that the meaning of openness is being settled not by states but in license files, largely outside the public eye.