Does the Government Have to Approve AI Models? No.

OpenAI showed off its next model family on August 1. One claim spread fast: Astra must pass a federal review before it can ship. Some reports said models now get "submitted to the federal government before public release."

That gets it backwards. There is a new federal review for advanced AI. It is opt-in. The order that made it rules out approval powers in plain words. If you want to guess when the next big model lands, that gap matters.

Key Takeaways

• Executive Order 14409 sets up an opt-in review for frontier AI models. It bars any required license or sign-off.
• Labs that join may share a model with agencies for up to 30 days before launch. They still pick the date.
• The order adds an NSA test for cyber skills and a new security clearinghouse.

What Executive Order 14409 actually says

President Trump signed the order on June 2, 2026. Its title is "Promoting Advanced Artificial Intelligence Innovation and Security."

The text is direct. It says nothing in the order "shall be construed to authorize creation of any mandatory governmental licensing, pre-clearance, or permitting requirement for the development, publication, release or distribution of AI models" (The White House, June 2026). The full text also sits in the Federal Register.

Law firms read it the same way. The order "does not mandate licensing or government approval of new models." Labs "may, on a voluntary basis, share covered frontier models with the federal government in advance of public release." It "stops short of imposing licensing requirements, mandatory safety testing or any government veto over launch decisions" (WilmerHale).

A late draft tightened this. The window shrank to 30 days. Drafters also made clear that joining cannot be tied to a sign-off (AI Governance Institute).

So there is no gate. There is a door, and walking through it is a choice.

Three terms do a lot of work here, and they get mixed up often:

• A covered frontier model is one the NSA test flags as strong enough to matter for cyber risk. That flag is the trigger for everything else.
• Prerelease access is a window of up to 30 days. Agencies study a model before the public sees it. It is not a review board.
• Pre-clearance means a rule that blocks a launch until an agency says yes. The order bans it outright.

Where the approval story came from

The mix-up is easy to understand. Two parts of the order do sound like enforcement.

First, there is a secret benchmark. The order asks for a way to test models for cyber skills and to flag covered frontier models (TechTarget). An NSA test sounds like a gate, even when it cannot stop a launch.

Second, the Astra coverage squeezed things. One report described a plan "requiring AI models to be submitted to the federal government before public release" (the-decoder). Squeeze a voluntary 30-day window into one clause and it reads like a rule.

There is a third reason. Many people expected licensing. It was the loudest idea in AI policy for two years. So when a review showed up, readers filed it under the thing they had argued about.

Here is the split, side by side.

The order does	The order does not
Create an opt-in prerelease framework	Require any model to join it
Allow up to 30 days of agency access	Give any agency a veto on launch
Direct a classified NSA cyber benchmark	Set a safety test a model must pass
Set up a cybersecurity clearinghouse	Create licensing or pre-clearance

What Astra is actually going through

OpenAI announced Astra on August 1, 2026. It skipped the usual press release. Instead it published a report on ten unsolved problems in math and computer science that an internal version had cracked.

Sam Altman has since shown the model to lawmakers. Those meetings are not the review. They are normal lobbying by a firm with a lot at stake in Washington.

Say OpenAI does join with Astra. Agencies would get access for up to 30 days. They would work with OpenAI on security tests. Then OpenAI picks the ship date, just as it would have anyway.

The company has not set a release date. It has not said whether Astra ships as GPT-6 or as a GPT-5 variant. Neither answer waits on Washington.

Why "voluntary" still bends the market

None of this makes the order weak. Reading "voluntary" as "irrelevant" would be a mistake.

A voluntary program sets a norm. Once one big lab joins, staying out is something a firm has to explain. It has to explain it to regulators, to buyers, and to anyone drafting purchase rules. Analysts note such programs can grow into formal rules later. They also note the programs favor big labs. Those labs already keep the lawyers, cleared staff and security teams a review demands.

That is the real story, and it matters more than a veto would. A rule that blocks launches hits everyone the same way. A program that rewards the staff to run a 30-day secret review favors whoever already has cleared people. The barrier is not the model. It is the paperwork.

The security half works the same way. A clearinghouse set up in July runs AI bug scans, checks what they find, and pushes patches to sectors such as hospitals, banks and utilities. That plan assumes labs already run strong models against real code. They do. Google's bug hunt turned up over a thousand Chrome flaws, which we covered in our piece on the Chrome security update AI results. Policy is catching up to practice.

What this means if you build on these models

For most builders, nothing changes about release planning. Ignore anyone who says a model is "awaiting approval."

Three things are worth tracking instead.

1. Designation, not approval. The live question is whether the NSA test flags a model as covered. That decides whether any of this applies.
2. A possible 30-day slip. If a lab joins, the window is a fair reason for a date to move by about a month. That is timing, not blocking.
3. Buyer questions. Expect vendor forms to start asking whether a lab took part. Voluntary programs gain force through buyers, not through the agency that wrote them.

For example, a hospital buying an AI coding tool in 2027 may well ask if the vendor's model went through the review. Nothing requires that question. It will get asked anyway.

Frequently Asked Questions

Does the government have to approve AI models before release?

No. Executive Order 14409 creates an opt-in review. It bars any required license or sign-off before a model ships. Labs keep control of timing.

Is OpenAI required to submit Astra for federal review?

No. Astra may be among the first models to use the plan, but joining is a choice. OpenAI has not given a date or said whether it ships as GPT-6.

What does the government get during the review?

Up to 30 days of early access to a covered frontier model, plus joint work on security tests. The NSA runs a separate secret benchmark to flag covered models.

Could this become mandatory later?

It could. Analysts say voluntary programs often become the base for formal rules. Nothing in this order does that. Its text blocks tying the review to a sign-off.

The short version

A federal review for frontier AI exists. It is real, and it is opt-in. No agency can stop OpenAI, Anthropic or Google from shipping a model.

If a launch slips, look at the usual causes first. Test results, capacity and rival timing move dates far more often. The better question is not whether Washington can block a model. It is what happens when an opt-in program becomes the price of selling to regulated buyers.