Washington now signs off on frontier models before you get them
GPT-5.6 is the first frontier model released after a US government review. Three weeks after the Fable 5 shutdown, the launch date of every major model runs through Washington, and Europe is about to add a checkpoint of its own.
OpenAI released GPT-5.6 to the public on July 9, after weeks in which only organisations on a government-managed access list could use it. The Commerce Department’s Center for AI Standards and Innovation tested the model, and senior officials had a say in when it could go out to everyone. GPT-5.6 is a solid step up in coding and agent work. For companies planning around new models, the review process matters more than the benchmarks, because every future frontier model is likely to go through it.
A launch with a permission slip, three weeks after a shutdown
On June 12 the Commerce Department ordered Anthropic to cut off foreign access to Claude Fable 5, and Anthropic ended up pulling the model for everyone for 19 days. I wrote about that episode when it ended. The concern was cyber capability: researchers had coaxed Fable 5 into finding and demonstrating software vulnerabilities.
GPT-5.6 raised the same concern, and this time the government stepped in before launch. According to CNBC’s report on the clearance, the administration asked OpenAI to stagger the release because of what the model could do. OpenAI called the process a collaborative back and forth, made a series of changes, and sent technical staff to Washington to answer questions. Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent and National Cyber Director Sean Cairncross were involved in the review.
The legal basis is an executive order that asks AI companies to give the government up to 30 days of access, voluntarily, before releasing models with advanced cyber capabilities. After watching Anthropic lose its flagship model for almost three weeks, no lab will treat that request as optional.
A licensing regime nobody voted for
In June the government showed it could switch off a deployed model overnight. In July it held one back until it was satisfied. No law was passed, but the most capable models from the two biggest US labs now reach the market on a schedule the US government helps set.
The case for this is reasonable. A model that can find exploitable bugs at scale is a security problem, and testing it before hundreds of millions of people have it makes more sense than recalling it afterwards. OpenAI got through in weeks, faster than many in the industry expected.
The costs fall on everyone else. Review adds time, and nobody can predict how much, because the criteria are not published. Staged access lists mean some organisations, mostly American, get a model first. And the process favours labs with the staff and political connections to handle it, which most of OpenAI’s and Anthropic’s challengers do not have.
The view from Copenhagen
European businesses have no seat in this process. The Fable 5 directive targeted foreign nationals, which from Denmark meant every user. A rollout that starts with vetted US organisations puts European customers at the back of the queue by default.
Europe is about to add a checkpoint of its own. From 2 August, the European Commission’s AI Office gets enforcement powers over providers of the most advanced general-purpose models. It can request information, demand access to models for evaluation, require risk mitigation and, as a last resort, order a model restricted, withdrawn or recalled from the EU market. The Commission says it prefers dialogue to enforcement. Even so, both sides of the Atlantic will soon have a way to stop a model.
What to change in your planning
Stop treating a launch announcement as an availability date. If a product plan depends on a model that has not shipped, build in a gap between announcement and production use, and assume European access comes later than American.
Keep a second model wired in. The businesses that got through the Fable 5 blackout switched models the same day. The same fallback protects you when a model you planned around is held up in review.
Follow the review criteria as closely as the benchmarks. Cyber capability triggers review today. Biology and autonomous agents are the likely next triggers, and the model that is best at the work you care about is also the one most likely to be gated.
New models will keep arriving quickly. From now on, each one reaches you through a government review first, and your plans should allow for the wait.
