Analysis
The Trump administration's new framework for reviewing advanced AI models before release will apply only to closed-source frontier systems, explicitly excluding open-weight models -- a carve-out that Axios reports was deliberate and is already reshaping how the closed-versus-open debate plays out in Washington.
Under the framework, a 'covered frontier model' is defined as closed-source, state-of-the-art in capability, and carrying meaningful national security risk. That definition puts OpenAI, Anthropic and Google's flagship models squarely inside the review process, while Meta's Llama models and other open-weight systems fall outside it entirely, with the framework explicitly stating nothing in it should be read as restricting open models once released.
“Under the framework, a 'covered frontier model' is defined as closed-source, state-of-the-art in capability, and carrying meaningful national security risk.”
The strategic logic, according to people familiar with the plan, is competitive: restricting open-source AI development domestically would hand ground to Chinese open-weight labs that are already racing to lead that specific category globally. The administration also isn't making the full framework public, leaving developers and outside researchers with limited visibility into exactly how pre-release review decisions will actually get made.
The practical effect, as one AI policy expert put it, is that the voluntary framework concentrates regulatory uncertainty onto a handful of the largest closed-model labs rather than spreading it across the entire AI ecosystem -- a meaningfully different approach than the EU's AI Act, which applies obligations more broadly regardless of open or closed licensing.
What to watch: whether Meta and other open-weight developers use the exemption to accelerate open model releases specifically because they fall outside the review framework, and whether closed-model labs push back on carrying disproportionate regulatory burden relative to open competitors building on the same underlying capability frontier.
