The newest AI models are no longer shipping the moment they are ready. On June 25, 2026, OpenAI agreed to release GPT-5.6 only to a small group of government-approved customers first, after the White House asked it to hold back a broad public launch. Two weeks earlier, the US government forced Anthropic to pull two of its most advanced models offline entirely. These AI model release delays put Washington, for the first time, between you and the next frontier model.
This shift started with a single executive order and a growing fear inside government about what these models can do on their own, especially in cybersecurity. Below we break down what is actually happening, which models are affected, why the government wants a first look, and what slower releases mean for anyone who relies on AI tools. We also clear up a common mix-up about whether Gemini 3.5 Pro is caught in the same net.
The Key Takeaways
- OpenAI agreed to stagger GPT-5.6, releasing it only to government-approved customers during a preview period before a broader launch “a couple of weeks later.”
- A June 2, 2026 executive order asks AI companies to voluntarily share frontier models with the government for up to 30 days of cybersecurity review before release.
- Anthropic was forced to take Fable 5 and Mythos 5 fully offline on June 12 under a Commerce Department export-control order, the harder precedent behind the OpenAI request.
- The core worry is autonomous cyber capabilities, models that can run multi-step attacks and find software vulnerabilities without a human guiding them.
- Gemini 3.5 Pro’s slip to July 2026 is officially about internal testing, not government review, though it lands in the same cautious climate.
What Is Actually Happening
For most of the AI race, the playbook was simple. A lab finished a model, ran its own safety checks, and shipped it to everyone at once. That pattern broke in June 2026. The US government is now inserting itself before the public release, and the biggest labs are going along with it.
The change is visible across two companies. OpenAI agreed to limit who gets GPT-5.6 at launch, with federal officials approving customers one at a time. Anthropic went further, pulling Fable 5 and Mythos 5 offline within hours of a government order. Different mechanisms, same direction. Both mean slower, gated releases with a government checkpoint in the middle.
What makes this strange is that there is no settled rulebook. The requests come from different parts of government, the review standards are still being written, and participation is technically voluntary. The result is a moment where the most powerful models in the world ship on the government’s timeline, not the lab’s.
Why the Government Wants to Review AI Models First
The trigger is cybersecurity, not chatbots saying the wrong thing. Officials in Washington and on Wall Street grew alarmed at what the latest models can do without a human in the loop. The specific concern is autonomous cyber capability, the reported ability to navigate multi-step attacks and pinpoint software vulnerabilities on their own.
Anthropic’s Mythos 5 was the model that set off the alarm. Its autonomous security skills were seen as powerful enough to create national-security risk if the wrong actor got access. When the government concluded that OpenAI’s GPT-5.6 was “on par” with Mythos on these capabilities, the same logic applied to OpenAI.
The fear is straightforward. A model that can find and exploit software flaws by itself is a defensive tool in the right hands and a weapon in the wrong ones. The government’s argument is that it should get to assess that power before a model reaches millions of users, including foreign adversaries.
The Executive Order Behind the Delays
The legal backbone is an executive order signed on June 2, 2026, titled “Promoting Advanced Artificial Intelligence Innovation and Security.” It sets up a voluntary framework where developers give the government early access to advanced models for up to 30 days before a wider release.
The review targets what the order calls a “covered frontier model.” Within 60 days, agencies including the NSA and CISA are tasked with building a classified benchmarking process to decide which models qualify based on their cyber capabilities. A May draft of the order had set the review window at 90 days, so the final 30-day figure is actually the lighter version.
Two things matter here. First, the order is voluntary, with no statute forcing a lab to comply. Second, it is only one of several levers, since the harder action against Anthropic came through export-control law, not this order. That overlap is why the rules feel confusing even to the companies involved.
ChatGPT 5.6: The First Model Held Back This Way
The OpenAI story, first reported by The Information on June 25, 2026, is the clearest example of the new normal. After a request from the White House, OpenAI agreed not to launch GPT-5.6 to everyone at once. Instead, federal leaders will approve access customer by customer during a preview period, with a broader rollout expected a couple of weeks later.
The request did not come from one tidy source. It originated with the Office of the National Cyber Director and the Office of Science and Technology Policy, while Commerce Secretary Howard Lutnick separately advised OpenAI against a broad launch without cross-agency sign-off. Sam Altman, OpenAI’s CEO, told staff the company was cooperating but made clear it was not happy about the setup.
“We’ve made clear to the US government that this is not our preferred long-term model, and will work with them and others in industry to achieve a more sustainable approach for future releases.”
For users, that means GPT-5.6 exists but you may not be able to touch it for weeks. A vetted enterprise or government partner gets it first, and general access follows once officials are satisfied.
Anthropic’s Fable 5 and Mythos 5: The Harder Precedent
OpenAI’s staggered launch looks gentle next to what happened to Anthropic. On June 12, 2026, the Commerce Department’s Bureau of Industry and Security, under Lutnick’s signature, ordered Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national. That order reached even Anthropic’s own foreign-national employees, inside or outside the US.
Anthropic disabled both models globally within hours, just days after launching them. This was not a request to slow down or stage a rollout. It was an export-control directive that pulled live frontier models off the market entirely, and it remains the most aggressive government action against a US AI lab to date. You can read the full sequence in our coverage of how Claude Fable 5 was pulled by the US government and our explainer on what Claude Mythos actually is.
The Anthropic case is the stick standing behind the OpenAI carrot. When the government asked OpenAI to limit GPT-5.6, both sides knew what a non-cooperative outcome could look like.
Is Gemini 3.5 Pro Delayed by the Government Too?
This is the common mix-up, so it is worth being precise. Gemini 3.5 Pro has slipped from a June to a July 2026 release, and it is easy to assume the same government review is the cause. Based on the public reporting, it is not.
Google’s stated reason is internal testing and performance refinement, with the model still in limited enterprise preview while the team improves coding, token efficiency, and long-task handling. Google declined to comment on the revised schedule, and no report ties the delay to a federal review. So while Gemini 3.5 Pro is launching slower, the cause appears to be Google’s own bar, not Washington’s.
The nuance still matters for the bigger picture. Even delays that are purely internal now happen against a backdrop where the government is actively gating other frontier models. Caution is becoming the default mode across the industry, whether the pressure comes from regulators or from the labs themselves.
How Each Model Is Affected
| Model | Company | Status | Driven by | Date |
|---|---|---|---|---|
| GPT-5.6 | OpenAI | Staggered release, government-approved customers first | White House request (ONCD + OSTP) | June 25, 2026 |
| Fable 5 | Anthropic | Pulled offline globally | Commerce export-control order | June 12, 2026 |
| Mythos 5 | Anthropic | Pulled offline globally | Commerce export-control order | June 12, 2026 |
| Gemini 3.5 Pro | Delayed to July | Internal testing (not government) | June 2026 |
The table makes the split clear. Three of the four delays trace back to government action, while Gemini’s is a company decision. Together they show how fast the release environment has changed in a single month.
What Slower AI Model Releases Mean for You
The most immediate effect is access. The newest, most capable models may now reach vetted partners and government-approved customers before they reach the general public, so the gap between “announced” and “available to me” is widening. If you build on these APIs, you may face waitlists or staged access you did not see a year ago.
There is also a fragmentation risk. Some powerful models could end up restricted by nationality or region, as the Anthropic order shows, which complicates global teams and any product that serves international users. Multi-model tools that route across providers, like Fello AI, help here by letting you fall back to whatever model is actually available to you rather than betting on a single lab.
The upside is harder to see but real. If the reviews work as intended, the models that do reach you have been checked for the worst cyber-abuse cases first. The open question is whether a voluntary, multi-agency process can stay fast enough to keep US releases competitive, a tension that mirrors the debate around the EU AI Act and how rules shape which tools you can use.
Conclusion
The era of instant frontier-model launches is over, at least in the US. A June 2026 executive order, an export-control order that pulled Anthropic’s top models offline, and a staggered GPT-5.6 release have together put the government in the release path for the most capable AI. Expect more previews, more vetted-customer rollouts, and a longer wait between announcement and access.
For now, the practical move is to stay flexible. Do not build a workflow that depends on one model being available everywhere on day one, and keep a backup provider ready. The competitive picture between the big labs is shifting fast, and our breakdown of Anthropic vs OpenAI is a good place to see who is best positioned as the rules tighten.
FAQ
Why is the US government reviewing AI models before release?
The main concern is autonomous cybersecurity capability. Officials worry that frontier models can run multi-step cyberattacks and find software vulnerabilities without human help, so they want to assess that power before a model reaches the public or foreign adversaries.
Is ChatGPT 5.6 delayed?
Not fully delayed, but staggered. After a White House request reported on June 25, 2026, OpenAI agreed to release GPT-5.6 to government-approved customers first, with a broader public rollout expected a couple of weeks after the preview period.
Why were Anthropic’s Fable 5 and Mythos 5 pulled offline?
On June 12, 2026, the Commerce Department issued an export-control order barring foreign nationals from accessing the two models, citing their advanced autonomous cyber capabilities. Anthropic disabled both globally within hours, just days after launch.
Is Gemini 3.5 Pro delayed by the government?
No. Gemini 3.5 Pro slipped from June to July 2026, but Google’s stated reason is internal testing and performance refinement, not a government review. No public report links its delay to federal oversight.
Is the government review mandatory for AI companies?
The June 2, 2026 executive order is voluntary and asks for up to 30 days of pre-release access to “covered frontier models.” The Anthropic action, however, came through binding export-control law, so the overall picture mixes voluntary requests with enforceable orders.
