U.S. lifts restrictions on OpenAI's GPT-5.6 after negotiations with the Trump administration

The Axios article (by Ashley Gold and Ina Fried) reveals that the Trump administration has given the green light to the broad launch of GPT-5.6, OpenAI's advanced model, according to a source who confirmed it to Axios. OpenAI announced that its flagship GPT-5.6 model, called "Sol," along with the lower-tier versions "Terra" and "Luna," will be released publicly on Thursday.

The relevant point here is not merely technical but political: the U.S. government and the most advanced AI companies are negotiating case by case, in real time, how access is granted to technologies considered potentially dangerous. This marks an ad hoc regulatory pattern that is taking hold in 2026, without a clear and stable regulatory framework.

According to the article, the approval came after additional testing and meetings between OpenAI and government officials. The testing was carried out by the Center for AI Standards and Innovation, part of the Department of Commerce, with technical experts from OpenAI who remained in Washington, D.C. to answer questions.

The context is key: last month, the Trump administration had pressed OpenAI to carry out a staggered launch of GPT-5.6, limiting initial access to government-approved entities. OpenAI had stated at the time that such a staggered rollout was not its preferred method. The company also noted that AI companies and the government are operating without concrete standards for the launch of these models—standards that are supposed to be defined based on Trump's latest executive order on AI.

The broader picture shows that both OpenAI and Anthropic have recently had to limit the launch of their most powerful models. In June, the Department of Commerce barred foreigners from accessing Anthropic's Mythos and Fable models, essentially forcing their withdrawal from the market. The ban on Fable was lifted last week, with customer access restored a day later.

A particularly interesting point in the article is the official contradiction: a White House official denied that the administration had given a "green light," an approval, or an authorization, arguing that no such permission is necessary. According to that official, "no such permission is required or granted," and decisions about the timing and scope of launches "rest entirely with the companies." This official cited Trump's June 2 executive order, which prohibits any mandatory federal licensing or prior authorization for the launch of AI models, and stated that any testing or meetings with government experts are voluntary.

For Manuel's newsletter, this illustrates an underlying tension in the governance of agentic AI and frontier models: on one hand, there is an executive order that formally prohibits mandatory pre-authorization; on the other, in practice there is a de facto process of testing, meetings, and informal "green light" that conditions when and how the most powerful models are launched. This ambiguity between what the rule says and what actually happens in practice is a recurring theme to watch: the leading AI companies (OpenAI, Anthropic) appear to be subject to government scrutiny that is not formally codified but that nonetheless determines the pace of their commercial launches.

Note: the article is relatively brief and does not delve into the specific technical capabilities of GPT-5.6 beyond mentioning its three tiers (Sol, Terra, Luna); nor does it detail exactly what the safety testing conducted by the Center for AI Standards and Innovation involved.