Two weeks ago, the US government forced Anthropic to pull its most powerful model offline in the first-ever export control on an AI model itself. Today, it started letting it back. Anthropic announced that the government has cleared Claude Mythos 5, its strongest cybersecurity model, to be redeployed to a set of US organizations that operate and defend critical infrastructure. Reporting from Bloomberg, CNBC, and Reuters puts that set at roughly 100 companies and federal agencies. Commerce Secretary Howard Lutnick, who signed the June 12 order that started the standoff, wrote in a new letter that "appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model." It is a major reversal, and a careful one. This is not the ban being lifted. It is the government deciding exactly who is trusted enough to wield the most powerful vulnerability-hunting AI ever built, and handing it back to them and only them. Here is what was actually restored, what is still locked, and what this resolution really tells us.

The reversal is real but narrow, and the details are the whole story.

Mythos 5 is back, but only for about 100 specific organizations. These are US companies and federal agencies that operate and defend critical infrastructure, power, water, healthcare, communications, the systems where an attack could affect millions. This is essentially the Project Glasswing model restored: the most powerful AI is handed to the defenders who need it most, and withheld from everyone else.

The most striking detail is who is allowed back in. The original June 12 order was brutally broad, it banned access by any foreign national, anywhere, including Anthropic's own non-American employees. Today's clearance reportedly lifts that for the approved organizations, meaning non-American staff at those critical-infrastructure partners, and at Anthropic itself, can use Mythos 5 again. That was one of the most disruptive parts of the ban, and it is being walked back for trusted partners.

What is NOT restored: Fable 5, the model the public could actually use. It remains offline for general use. Anthropic said it is "continuing to work with the government to expand access to Mythos 5 and make Fable 5 available for general use again." So the everyday, generally-available frontier model from Anthropic is still dark. Today helped the defenders. It did not bring the public model back.

This resolution tells you what the standoff was really about, and it was never simply "Anthropic's model is dangerous."

Recall the arc. On June 12 the government pulled both models, citing national security and a jailbreak that could turn the safety-limited Fable 5 into the unrestricted Mythos 5. Days later, the concern deepened: an NSA red-team exercise reportedly showed Mythos breaking into nearly all of a set of classified systems, not in weeks but in hours. Then on Monday, the Five Eyes intelligence alliance warned that frontier AI could reshape the cyber threat landscape within months. The fear was concrete and specific: a model that finds and exploits critical vulnerabilities faster than any human, in the wrong hands.

Seen that way, today's outcome is perfectly logical. The same capability that makes Mythos dangerous in an enemy's hands makes it invaluable in a defender's. So the government did not choose between banning it and freeing it. It chose to control exactly who holds it, restoring it to the power plants, hospitals, and agencies defending the nation, while keeping it away from the open market. The resolution reveals the real policy: this class of capability is now treated like a controlled asset, distributed to vetted defenders, withheld from everyone else. The ban was never the end goal. Gated access was.

A few things are true at once here, and your skeptical eye should hold all of them.

This is a genuine win for Anthropic, and a fast one. Two weeks from an unprecedented global shutdown to a negotiated restoration is remarkably quick for anything involving the federal government, and it suggests both sides wanted a workable answer rather than a fight. Senior Anthropic staff reportedly flew to Washington to hammer it out, and the speed signals the government values the defensive capability too much to keep it switched off.

But the precedent it sets is the bigger story, and it cuts both ways. The government has now established that it can pull a frontier model off the market overnight, and then dictate the precise list of who gets it back. That is enormous power over a private company's flagship product, and it is now demonstrated, not theoretical. For Anthropic, racing toward an October IPO, this is a double-edged disclosure: it proves Mythos is powerful enough to matter to national security, while proving the company's most advanced products can be switched on and off by a Commerce Department letter. Both of those facts now live in the S-1.

And the resolution is partial. Fable 5, the model real customers built on, is still down. Anyone who wired Fable 5 into a product three weeks ago is still waiting, with only a promise that Anthropic is "working on it." The defenders got served first. The market is still in limbo.

The Mythos news did not happen in isolation. On the very same day, OpenAI launched GPT-5.6, and Washington gated it the same way it gated Anthropic.

OpenAI rolled out GPT-5.6 today, but is limiting access to it at the government's request. The model comes in three versions: Sol, the most powerful, Terra, balanced for efficiency and power, and Luna, built for speed and affordability, plus an "ultra" mode that splits work across multiple sub-agents. It is launching as a limited preview to only about 20 companies, each one individually approved by the government. The concern is identical to the one behind the Mythos ban: cybersecurity capability. OpenAI says it believes GPT-5.6 Sol "is better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks," and that its capabilities do not reach the "critical" level in its safety framework.

The significance is what this pairing reveals. Two weeks ago it was possible to read the Anthropic ban as Anthropic being singled out. Today proves it is policy. OpenAI is now negotiating the same safeguards, accepting the same customer-by-customer government approval, releasing its newest model on the same gated terms. OpenAI made its discomfort clear: "We don't believe this kind of government access process should become the long-term default." But it complied anyway. On a single day, the two leading AI labs both released their most powerful models only to a government-approved list. That is no longer an exception. That is the new system.

The US government reversed course today and cleared Anthropic's Mythos 5, its most powerful cybersecurity model, for redeployment to roughly 100 US organizations that operate and defend critical infrastructure, two weeks after forcing it offline in the first-ever export control on an AI model. Commerce Secretary Howard Lutnick said appropriate safeguards now permit "certain trusted partners" to access it, and the clearance even restores access for non-American employees at those organizations who were swept up in the original ban. The catch: this is a narrow, gated restoration, not a lifted ban, and Fable 5, the publicly usable model, remains offline for general use while Anthropic keeps negotiating. And on the same day, OpenAI launched GPT-5.6, in three tiers (Sol, Terra, Luna), but only to about 20 government-approved companies, under the exact same kind of restrictions, proving Anthropic was never being singled out. The resolution reveals what this was always about: not banning powerful AI, but controlling exactly who holds it. It is a fast win for Anthropic, but it sets a heavy precedent weeks before its IPO, the government can pull a frontier model overnight and dictate who gets it back, and it now applies to OpenAI too. Two weeks ago the question was whether Washington would switch off a model. Today it answered a bigger one: who decides who gets the most powerful AI. For now, the answer is the government.

Stay building. 🤖