OpenAI argues the 30-day AI safety review shouldn't become permanent
The voluntary 30 day AI safety review let GPT 5.6 ship in about half the time. OpenAI says that argues against making the review the long term default.
The voluntary 30 day AI safety review let GPT 5.6 ship in about half the time. OpenAI says that argues against making the review the long term default.
President Trump signed an executive order in early June asking frontier AI labs to give the federal government 30 days of pre-release review on their most powerful models. OpenAI's GPT-5.6 family ships to the public on July 9, having been with a small group of trusted partners for less than half that window, and the company has said the experience argues against making the framework the long-term default.
The model release is the easy part. The harder question is whether the 30-day review framework created by Executive Order 14409, titled "Promoting Advanced Artificial Intelligence Innovation and Security," is a real gate on frontier AI deployment or a courtesy that well-positioned companies can negotiate down.
The order, signed June 5, asks AI developers to voluntarily submit their most capable systems to the federal government a month before public release. It also tasks the Office of the National Cyber Director with building a vulnerability-sharing channel so cyber findings on frontier models can reach critical-infrastructure operators (banks, utilities, hospitals) before those models go live. It is the Trump administration's central answer to the question of how Washington keeps a hand on the most consequential AI systems without legislating. The text is voluntary. Compliance is, on paper, a request.
The company first showed GPT-5.6 Sol to a "small group of trusted partners" in late June at the administration's request, and used its preview page and the GPT-5.6 Preview System Card to disclose what was being shown. According to Axios reporting cited by Channel News Asia, the Trump administration then gave OpenAI approval to roll the model out to the public without completing the full 30-day window.
OpenAI's public posture is more revealing than its private dealings. The company said, "We don't believe this kind of government access process should become the long-term default," and said it was complying only because the alternative was worse for getting the new model family out the door. That is a frontier lab publicly signaling that the regime it is currently operating under is one it would not choose to extend. A permanent version of the same arrangement, in OpenAI's telling, would be a problem for the industry rather than a feature of it.
The precedent matters more than the launch. If a frontier model can reach the public in roughly two weeks because the federal government is willing to bless a shorter clock, the 30-day window is a ceiling, not a floor. It can be invoked or waived. EO 14409's text is voluntary, and the GPT-5.6 release is the first time that voluntariness has been visibly tested at the frontier. Subsequent releases, from OpenAI or its competitors, will be measured against a benchmark the order itself never set: how fast can a well-resourced lab and a sympathetic administration make the review go.
The other piece worth watching is what gets shared during the truncated review. The order's core deliverable is supposed to be a vulnerability channel to critical infrastructure. If the cyber findings from GPT-5.6's pre-release review reach banks, utilities, and hospitals before the public launch, the mechanism is functioning as designed. If they don't, the order has been reduced to a notification, and OpenAI's characterization of the process as something that should not become permanent starts to look less like principled resistance and more like an accurate read of the terrain.
For now, the takeaway is narrow but concrete. The Trump administration has its first real data point on whether the voluntary framework functions as a gate, and the data point is that it depends on who is asking. OpenAI has a new model family shipping on July 9, with three variants (Sol, Luna, and Terra) for different use cases. The open question of whether the next frontier release, from OpenAI or anyone else, will face the same compressed review is a question the order itself does not answer.