White House to OpenAI and Anthropic: U.S. Review Comes Before British Testing

The White House has told OpenAI and Anthropic to let American officials review new models before those models are handed to British testers. The directive, dated September 25, 2026, marks a turning point in how frontier AI gets checked, shared, and governed. It is a small sentence with enormous consequences, because it quietly rewrites the order of operations for the most powerful technology of our time.

Until now, the story of AI safety testing has been one of openness. Labs built models, shared them with trusted outside evaluators, and let a global community of researchers poke at them. That era is now being replaced by something more guarded: a sequence in which the United States gets the first look, and allies get the second.

What the Directive Actually Changes

The mechanics are simple to describe. When OpenAI or Anthropic finishes a new frontier model, the company is now expected to route it through a U.S. government review first. Only after that review is complete should the model be shared with British testers, the outside evaluators who examine models for dangerous capabilities, misuse risks, and unexpected behavior.

What makes this significant is not the review itself. Governments have been examining AI systems for years. What makes it significant is the ordering. The United States is now formally placing itself first in the queue of who gets to understand a new model, ahead of a close ally with its own sophisticated testing capacity.

That ordering is a claim about sovereignty. It says that American-developed frontier AI is, in the eyes of Washington, a national asset before it is a global one. And it says that the first and most informed look at a new model belongs to the country where it was built.

Why Washington Wants the First Look

Three forces appear to be driving this. Understanding them helps explain why this is likely a permanent shift rather than a one-off request.

1. Security concerns have moved from theory to policy

Frontier models are now capable enough that their capabilities are treated as strategically relevant. A model that can reason through complex scientific problems, write and debug large amounts of code, or coordinate multi-step tasks is not just a product, it is a capability with national security dimensions. If a foreign government, researcher, or bad actor sees those capabilities before the U.S. government does, Washington loses the ability to prepare.

2. Shared testing creates shared exposure

Every time a model is shared with an outside tester, information moves. Those testers have their own systems, their own staff, and their own vulnerabilities. From a security standpoint, each handoff is a point of potential leakage. Requiring a U.S. review first creates a checkpoint where officials can assess what is actually being handed over, and to whom.

3. The U.S. wants to set the terms

Whoever tests first sets the vocabulary. The first evaluation framework tends to become the benchmark everyone else reacts to. By placing itself first, the U.S. government is positioning itself to define what "safe enough to share" means, rather than reacting to definitions created elsewhere.

Why the British Testing Relationship Matters So Much

The mention of British testers is not incidental. It signals that the U.S.-UK AI relationship is being restructured, not ended. British evaluators have built a reputation for serious, technically deep model testing. They are not a formality in this process, they are a genuine second opinion.

But a second opinion only works if it is independent. That is the tension at the heart of this directive. If British testers only see models after U.S. officials have signed off, are they still an independent check? Or are they a confirmatory step, useful, but no longer capable of catching something the first reviewer missed?

There is a real risk here. Independent review is valuable because reviewers think differently, look for different failure modes, and bring different cultural and regulatory assumptions. If the U.S. review becomes a gate that shapes what British testers even receive, that diversity of perspective narrows. The process becomes more controlled, but possibly less effective at finding the unexpected.

The counterargument is equally strong. A model that has been reviewed once is a known quantity. Sharing a reviewed model with allies is safer than sharing an unreviewed one. And in a world where AI capabilities are advancing faster than governance, some ordering is better than none.

The New Geopolitics of Model Access

Step back, and this directive looks like the beginning of an access hierarchy. Frontier AI is becoming a controlled resource, and access to it is being tiered.

• Tier one: the developing company and the U.S. government.
• Tier two: trusted allies and their testing bodies, after review.
• Tier three: enterprise customers, under contract and monitoring.
• Tier four: the general public, through consumer products with guardrails.

This is not a new pattern. It mirrors how the world handles nuclear technology, advanced semiconductors, and certain classes of cryptography. What is new is that it is being applied to software that many companies already build their products on top of, and applied before the rules are fully written.

The long-term implication is significant. If frontier AI access becomes a matter of diplomatic relationships, then the countries with the closest ties to the leading labs will get the best technology first. Others will get it later, in a more limited form, or not at all. That reshapes competitive advantage at a national level, not just a company level.

What This Means for AI Developers

For OpenAI, Anthropic, and anyone else building at the frontier, the practical effects are concrete.

Release timelines get longer and less predictable

A review step inserts a variable into launch planning that companies do not control. When you announce a model, you want to know when it ships. When a government is in the loop, that certainty softens. Expect product roadmaps to build in buffer time they did not need before.

Compliance becomes a competitive function

Labs that build strong relationships with regulators, document their safety work well, and can move quickly through review will release faster than those that cannot. Policy and safety teams shift from cost centers to speed enablers. This is a meaningful change in how AI companies will organize themselves.

Smaller labs face a harder path

Large labs have the legal, policy, and security staff to handle government review. Smaller startups often do not. If review becomes standard for frontier models, it becomes another barrier that favors incumbents. That is a real risk for the health of the AI ecosystem, and something policymakers will need to address directly, or the frontier will consolidate into a handful of players.

What This Means for Businesses

If you run a company that depends on frontier AI, this directive affects you more than it first appears.

• Plan for model version lag. New flagship models may reach you later than they would have. Build flexibility into any product that depends on a specific model's capabilities.
• Expect tiered access. Some capabilities may be available to large, vetted enterprise customers before they reach broader markets. If a capability is critical to your product, start the conversation early.
• Treat provider stability as a real risk. Regulatory friction can slow a provider's releases even if the technology is ready. Diversifying across providers, or keeping a self-hosted option in reserve, is a sensible hedge.
• Watch where your data flows. As models move through review processes, questions about data handling and jurisdiction will grow. Understand where your prompts and outputs travel.

The businesses that adapt best will be the ones that stop treating AI model access as guaranteed and start treating it as a managed supply chain.

The Risks of Moving First and Testing Second

Not everyone will see this directive as a win. Three concerns deserve serious weight.

Slowdown risk. Every additional step between a finished model and a deployed one is time in which good applications, medical research, scientific discovery, accessibility tools, are delayed. Safety has a cost, and that cost is real.

Alliance friction. Telling a close ally to wait its turn is a diplomatic statement. If the U.S. claims first review over models developed on its soil, other countries may respond by claiming first review over models developed on theirs. The result could be a fragmented global testing regime where no one sees the whole picture.

False confidence. A U.S. review that happens first may create a feeling that the model has been cleared. But safety testing is never finished. Models behave differently in the wild than in a lab. A first review is a starting point, not a certificate.

What to Watch Next

Several signals will reveal how this plays out.

• Whether the review process gets a published standard, or stays case-by-case.
• Whether smaller AI developers are brought into the same requirement, or exempted.
• Whether British testers describe their access as genuine or constrained.
• Whether other countries adopt a similar "review before sharing" posture.
• Whether release timelines for frontier models visibly lengthen.

The answers to these questions will determine whether this directive becomes a thoughtful checkpoint or a slow-moving bottleneck. Right now, the direction is clear: the United States has decided that seeing a new AI model first is a matter of national interest.

The Bigger Picture

For most of AI's short history, the assumption was that capability moves faster than rules, and rules chase capability. This directive flips that assumption in one narrow but important way. It puts a rule in place before the next generation of models arrives, rather than after.

That is a meaningful shift. It suggests governments are no longer playing catch-up on AI, they are trying to get ahead of it. Whether they can move fast enough, and whether they can do it without choking off the openness that made AI advance so quickly, is the defining question of the next few years.

For builders, the message is to treat regulatory process as part of the product. For businesses, it is to treat model access as something to plan for, not something to assume. And for everyone else, it is a reminder that the most powerful software ever created is now being handled less like software, and more like a strategic resource.