The Government Just Got a Veto Over the Best AI. No One Voted on It.

On Friday, OpenAI released the most capable artificial intelligence models it has ever built. Almost no one is allowed to use them.

That is not a capacity problem. It is not a staggered rollout to manage server load. It is a deliberate restriction, made at the request of the United States government, and OpenAI agreed to it. The company's three new models - named Sol, Terra, and Luna - are being released first to roughly 20 "trusted partners," a list OpenAI shared with the federal government before launch. Everyone else waits.

The capability concerns behind this are real. The precedent it sets is the part that should keep you up at night.

What Actually Happened

OpenAI announced the GPT-5.6 family on Friday: Sol, its most powerful model yet and, by the company's own description, its strongest tool ever for cybersecurity, with notable gains in coding and biology; Terra, a balanced everyday model; and Luna, a faster, cheaper option.

Ahead of the launch, OpenAI previewed the models for the federal government and shared its rollout plans. The Trump administration asked the company to limit the release. OpenAI complied, restricting initial access to a small group of vetted partners - reportedly around 20, with Amazon's Bedrock platform serving as one route in - and providing that partner list to the government.

OpenAI did not pretend to be happy about it. In its own statement, the company said plainly: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." It framed the restriction as a "short-term step," with broader availability promised "within coming weeks."

Read that again. The company being restricted is the one warning you about the restriction.

The Legitimate Half of the Story

It would be easy, and wrong, to treat this as cartoonish government overreach. There is a serious argument on the other side, and honesty requires stating it clearly.

Frontier AI models have crossed into genuinely dual-use territory. When OpenAI says Sol is its most capable cybersecurity model and has improved at biology, those are not marketing flourishes - they are exactly the capabilities that national-security officials worry about. A model that can meaningfully accelerate vulnerability discovery, or lower the barrier to engineering a pathogen, is not the same kind of product as a better chatbot. The "uplift" such a system gives a malicious actor is the central question in AI safety research right now.

If you accept that premise - that some capabilities are dangerous enough to warrant caution about who gets them, and when - then a brief, deliberate pause before mass release is not insane. It may even be responsible. The question was never whether powerful AI should ever be subject to any limits.

The question is who decides, by what rule, and with what accountability. And on that question, what happened Friday should alarm you regardless of your politics.

There Is No Law Here

This is the heart of it. There is no statute that authorized this. No regulation. No congressional vote, no rulemaking, no court order, no published standard that any model is measured against. There was a request from the executive branch and a decision by a private company to honor it.

That means the government of the United States now holds an informal veto over who may access the most powerful AI in the world - and it is being administered by the vendor, off the books, with the approved-partner list handed to the state. A governance structure of real consequence was assembled by handshake, in a week, with no public input and no way to see the rules.

"Trusted partners" is a phrase worth sitting with. Trusted by whom? Vetted against what criteria? With what appeal if you are left off the list? When access to a transformative technology runs through a roster that the executive branch helps shape, the government is no longer just regulating a market. It is picking who is in it.

The Word to Watch Is "Repeatable"

OpenAI said it is working with the administration to build a framework for these assessments - tied to a coming executive order on cybersecurity - and to develop a "repeatable process for future model releases."

Repeatable. That is how a one-time exception becomes permanent infrastructure. The first time a government quietly gatekeeps a technology, it is an emergency measure. The second time, it is policy. By the third, it is simply how things work, and no one remembers that there was ever a moment to object. We are at the first time right now.

The same lever that can throttle bioweapon uplift can throttle a competitor, a foreign ally, a research lab, a newsroom, or a developer the administration of the day happens to dislike. Once the machinery exists - vendor restricts on request, hands over the access list, calls it a framework - the only thing standing between "narrow safety measure" and "discretionary control over who gets the future" is the goodwill of whoever happens to be in office. That is not a foundation. That is a hope.

What a Real Answer Looks Like

If frontier models are genuinely dangerous enough to ration - and they may be - then the decision to ration them is exactly the kind of decision a democracy is supposed to make in the open. That means law, not arrangement. It means transparent, published criteria for what triggers a restriction. It means due process for who qualifies as a "trusted partner," with a way to contest exclusion. It means oversight by someone other than the people issuing the requests, and sunset clauses so that "short-term" actually ends.

The answer to "should every developer get the bioweapon-capable model on day one?" might legitimately be no. But the answer to "who decides, and how do we hold them accountable?" can never be "the President asked, and the most powerful company in the industry said yes." One of those is a safety policy. The other is just power, wearing a safety policy's clothes.

The models are named for the sun, the earth, and the moon - Sol, Terra, Luna - which is the kind of grandeur the industry loves. The real story is smaller and far more consequential: in a single week, with no vote and no statute, the executive branch established a working veto over the most powerful technology of the decade, and the company on the receiving end is the only party on record saying it shouldn't last.

It will last exactly as long as the rest of us let it pass without comment.

Frequently Asked Questions

Did the U.S. government ban OpenAI's new models?
No. The government did not ban anything. The Trump administration requested that OpenAI limit the initial release of its new GPT-5.6 models, and OpenAI voluntarily complied by restricting early access to roughly 20 government-shared "trusted partners." There is no law or regulation compelling the restriction - it rests on a request and the company's decision to honor it.

What are OpenAI's Sol, Terra, and Luna models?
They are the three models in OpenAI's new GPT-5.6 family, announced in late June 2026. Sol is the flagship and, per OpenAI, its most capable model yet for cybersecurity, with gains in coding and biology. Terra is a balanced general-purpose model, and Luna is a faster, lower-cost option.

Why did the government want the release limited?
The stated concern is dual-use risk: frontier models capable in cybersecurity and biology can also lower the barrier for malicious actors. A short pause before broad release is meant to allow safety assessment. The dispute is not over whether such risks exist, but over the mechanism - an informal request rather than transparent, accountable law.

Does OpenAI support the restriction?
Only as a temporary step. OpenAI publicly stated it does not believe this kind of government access process "should become the long-term default," warning that it withholds the best tools from defenders, developers, and global partners. The company has promised broader availability within weeks.

Why does this matter if it's only temporary?
Because OpenAI and the administration are reportedly building a "repeatable process" for future releases, tied to a coming cybersecurity executive order. A one-time exception that becomes a standing framework hands the executive branch ongoing, discretionary influence over who can access frontier AI - without the transparency, criteria, or oversight that a formal law would require.